# M OSES Statistical Machine Translation

#### Document technical information

Format pdf
Size 1.7 MB
First found Jun 9, 2017

#### Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

#### Transcript

MOSES
Statistical
Machine Translation
System
User Manual and Code Guide
Philipp Koehn
[email protected]
University of Edinburgh
Abstract
This document serves as user manual and code guide for the Moses machine translation decoder. The
decoder was mainly developed by Hieu Hoang and Philipp Koehn at the University of Edinburgh and
extended during a Johns Hopkins University Summer Workshop and further developed under EuroMatrix and GALE project funding. The decoder (which is part of a complete statistical machine translation
toolkit) is the de facto benchmark for research in the field.
This document serves two purposes: a user manual for the functions of the Moses decoder and a code
guide for developers. In large parts, this manual is identical to documentation available at the official
Moses decoder web site http://www.statmt.org/. This document does not describe in depth the underlying
methods, which are described in the text book Statistical Machine Translation (Philipp Koehn, Cambridge
University Press, 2009).
October 14, 2014
2
Acknowledgments
The Moses decoder was supported by the European Framework 6 projects EuroMatrix, TC-Star,
the European Framework 7 projects EuroMatrixPlus, Let’s MT, META-NET and MosesCore
and the DARPA GALE project, as well as several universities such as the University of Edinburgh, the University of Maryland, ITC-irst, Massachusetts Institute of Technology, and others.
Contributors are too many to mention, but it is important to stress the substantial contributions
from Hieu Hoang, Chris Dyer, Josh Schroeder, Marcello Federico, Richard Zens, and Wade
Shen. Moses is an open source project under the guidance of Philipp Koehn.
Contents
1
2
Introduction
1.1 Welcome to Moses! .
1.2 Overview . . . . . .
1.2.1 Technology .
1.2.2 Components
1.2.3 Development
1.2.4 Moses in Use
1.2.5 History . . .
1.3 Get Involved . . . .
1.3.1 Mailing List .
1.3.2 Suggestions .
1.3.3 Development
1.3.4 Use . . . . . .
1.3.5 Contribute .
1.3.6 Projects . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Installation
2.1 Getting Started with Moses . . . . . . .
2.1.2 Basic Setup . . . . . . . . . . . .
2.1.3 Other software to install . . . . .
2.1.4 Platforms . . . . . . . . . . . . .
2.1.5 Windows Installation . . . . . .
2.1.6 OSX Installation . . . . . . . . .
2.1.7 Linux Installation . . . . . . . .
2.1.8 Run Moses for the first time . .
2.1.9 Chart Decoder . . . . . . . . . .
2.1.10 Next Steps . . . . . . . . . . . .
2.1.11 bjam options . . . . . . . . . . .
2.2 Building with Eclipse . . . . . . . . . .
2.3 Baseline System . . . . . . . . . . . . . .
2.3.1 Overview . . . . . . . . . . . . .
2.3.2 Installation . . . . . . . . . . . .
2.3.3 Corpus Preparation . . . . . . .
2.3.4 Language Model Training . . .
2.3.5 Training the Translation System
2.3.6 Tuning . . . . . . . . . . . . . . .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11
11
11
11
12
13
14
14
14
14
15
15
15
15
16
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
21
21
21
21
23
24
24
25
25
26
27
28
28
30
31
31
31
32
34
36
36
4
CONTENTS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Tutorials
3.1 Phrase-based Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 A Simple Translation Model . . . . . . . . . . . . . . . . . .
3.1.2 Running the Decoder . . . . . . . . . . . . . . . . . . . . . .
3.1.3 Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.4 Verbose . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.5 Tuning for Quality . . . . . . . . . . . . . . . . . . . . . . . .
3.1.6 Tuning for Speed . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.7 Limit on Distortion (Reordering) . . . . . . . . . . . . . . . .
3.2 Tutorial for Using Factored Models . . . . . . . . . . . . . . . . . .
3.2.1 Train an unfactored model . . . . . . . . . . . . . . . . . . .
3.2.2 Train a model with POS tags . . . . . . . . . . . . . . . . . .
3.2.3 Train a model with generation and translation steps . . . .
3.2.4 Train a morphological analysis and generation model . . .
3.2.5 Train a model with multiple decoding paths . . . . . . . . .
3.3 Syntax Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Tree-Based Models . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2 Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.3 Decoder Parameters . . . . . . . . . . . . . . . . . . . . . . .
3.3.4 Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.5 Using Meta-symbols in Non-terminal Symbols (e.g., CCG) .
3.3.6 Different Kinds of Syntax Models . . . . . . . . . . . . . . .
3.3.7 Format of text rule table . . . . . . . . . . . . . . . . . . . . .
3.4 Optimizing Moses . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.1 How much memory do I need during decoding? . . . . . .
3.4.2 How little memory can I get away with during decoding? .
3.4.3 Faster Training . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.4 Training Summary . . . . . . . . . . . . . . . . . . . . . . . .
3.4.5 Language Model . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.6 Suffix array . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.7 Cube Pruning . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.4.8 Minimizing memory during training . . . . . . . . . . . . .
3.4.9 Minimizing memory during decoding . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
53
. 53
. 53
. 54
. 55
. 56
. 60
. 61
. 64
. 65
. 66
. 67
. 70
. 71
. 72
. 73
. 73
. 75
. 79
. 80
. 84
. 85
. 89
. 91
. 91
. 93
. 94
. 96
. 97
. 99
. 99
. 100
. 100
2.4
2.5
3
2.3.7 Testing . . . . . . . . . . . . . . . . . . . .
2.3.8 Experiment Management System (EMS)
Releases . . . . . . . . . . . . . . . . . . . . . . .
2.4.1 Release 2.1.1 (3rd March, 2014) . . . . . .
2.4.2 Release 2.1 (21th Jan, 2014) . . . . . . . .
2.4.3 Release 1.0 (28th Jan, 2013) . . . . . . . .
2.4.4 Release 0.91 (12th October, 2012) . . . . .
2.4.5 Status 11th July, 2012 . . . . . . . . . . .
2.4.6 Status 13th August, 2010 . . . . . . . . .
2.4.7 Status 9th August, 2010 . . . . . . . . . .
2.4.8 Status 26th April, 2010 . . . . . . . . . . .
2.4.9 Status 1st April, 2010 . . . . . . . . . . .
2.4.10 Status 26th March, 2010 . . . . . . . . . .
Work in Progress . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
37
40
41
41
41
42
48
48
50
50
50
51
51
51
CONTENTS
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
101
102
102
103
104
107
110
115
123
User Guide
4.1 Support Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.2 Converting Pharaoh configuration files to Moses configuration files
4.1.3 Moses decoder in parallel . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.4 Filtering phrase tables for Moses . . . . . . . . . . . . . . . . . . . . .
4.1.5 Reducing and Extending the Number of Factors . . . . . . . . . . . .
4.1.6 Scoring translations with BLEU . . . . . . . . . . . . . . . . . . . . .
4.1.7 Missing and Extra N-Grams . . . . . . . . . . . . . . . . . . . . . . .
4.1.8 Making a Full Local Clone of Moses Model + ini File . . . . . . . . .
4.1.9 Absolutizing Paths in moses.ini . . . . . . . . . . . . . . . . . . . . .
4.1.10 Printing Statistics about Model Components . . . . . . . . . . . . . .
4.1.11 Recaser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.12 Truecaser . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.13 Searchgraph to DOT . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1.14 Threshold Pruning of Phrase Table . . . . . . . . . . . . . . . . . . .
4.2 External Tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Word Alignment Tools . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.2 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.3 Part-of-Speech Taggers . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.4 Syntactic Parsers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.5 Other Open Source Machine Translation Systems . . . . . . . . . . .
4.2.6 Other Translation Tools . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Advanced Features of the Decoder . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Lexicalized Reordering Models . . . . . . . . . . . . . . . . . . . . .
4.3.2 Operation Sequence Model (OSM) . . . . . . . . . . . . . . . . . . . .
4.3.3 Class-based Models . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.4 Unsupervised Transliteration Model . . . . . . . . . . . . . . . . . .
4.3.5 Binary Phrase Tables with On-demand Loading . . . . . . . . . . . .
4.3.6 Binary Reordering Tables with On-demand Loading . . . . . . . . .
4.3.7 Compact Phrase Table . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.8 Compact Lexical Reordering Table . . . . . . . . . . . . . . . . . . . .
4.3.9 XML Markup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.10 Generating n-Best Lists . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.11 Word-to-word alignment . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.12 Minimum Bayes Risk Decoding . . . . . . . . . . . . . . . . . . . . .
4.3.13 Lattice MBR and Consensus Decoding . . . . . . . . . . . . . . . . .
4.3.14 Handling Unknown Words . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
127
127
127
127
127
128
128
129
129
129
130
130
131
131
132
132
134
134
136
137
138
140
140
141
141
144
147
149
151
155
155
157
158
159
160
162
163
164
3.5
4
3.4.10 Phrase-table types . . . . .
Experiment Management System .
3.5.1 Introduction . . . . . . . .
3.5.2 Requirements . . . . . . . .
3.5.3 Quick Start . . . . . . . . .
3.5.4 More Examples . . . . . . .
3.5.5 Try a Few More Things . .
3.5.6 A Short Manual . . . . . .
3.5.7 Analysis . . . . . . . . . . .
5
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
CONTENTS
4.3.15
4.3.16
4.3.17
4.3.18
4.3.19
4.3.20
4.3.21
4.3.22
4.3.23
4.3.24
4.3.25
4.3.26
4.3.27
4.3.28
4.3.29
4.3.30
4.3.31
4.3.32
4.3.33
4.3.34
4.4
4.5
5
Output Search Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Early Discarding of Hypotheses . . . . . . . . . . . . . . . . . . . . . . . .
Maintaining stack diversity . . . . . . . . . . . . . . . . . . . . . . . . . . .
Cube Pruning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Specifying Reordering Constraints . . . . . . . . . . . . . . . . . . . . . . .
Multiple Translation Tables and Back-off Models . . . . . . . . . . . . . .
Pruning the Translation Table . . . . . . . . . . . . . . . . . . . . . . . . .
Pruning the Phrase Table based on Relative Entropy . . . . . . . . . . . .
Multi-threaded Moses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Moses Server . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Continue Partial Translation . . . . . . . . . . . . . . . . . . . . . . . . . .
Global Lexicon Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Incremental Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Distributed Language Model . . . . . . . . . . . . . . . . . . . . . . . . . .
Suffix Arrays for Hierarchical Models . . . . . . . . . . . . . . . . . . . . .
Fuzzy Match Rule Table for Hierachical Models . . . . . . . . . . . . . . .
Translation Model Combination . . . . . . . . . . . . . . . . . . . . . . . .
Online Translation Model Combination (Multimodel phrase table type) .
Alternate Weight Settings . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Open Machine Translation Core (OMTC) - A proposed machine translation system standard . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.35 Pipeline Creation Language (PCL) . . . . . . . . . . . . . . . . . . . . . . .
4.3.36 Placeholder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.37 Modified Moore-Lewis Filtering . . . . . . . . . . . . . . . . . . . . . . . .
4.3.38 Constrained Decoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sparse Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.1 Word Translation Features . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.2 Phrase Length Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.3 Domain Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.4 Count Bin Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.5 Bigram Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4.6 Soft Matching Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Translating Web pages with Moses . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5.2 Detailed setup instructions . . . . . . . . . . . . . . . . . . . . . . . . . . .
Training Manual
5.1 Training . . . . . . . . . . . . . . . . . . .
5.1.1 Training process . . . . . . . . . .
5.1.2 Running the training script . . . .
5.2 Preparing Training Data . . . . . . . . . .
5.2.1 Training data for factored models
5.2.2 Cleaning the corpus . . . . . . . .
5.3 Factored Training . . . . . . . . . . . . . .
5.3.1 Translation factors . . . . . . . . .
5.3.2 Reordering factors . . . . . . . . .
5.3.3 Generation factors . . . . . . . . .
5.3.4 Decoding steps . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
164
166
166
167
168
169
170
172
175
175
177
178
179
182
185
187
188
189
191
193
194
197
200
201
201
202
203
204
205
205
205
206
206
209
213
213
213
214
214
215
215
216
216
217
217
217
CONTENTS
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
5.13
5.14
6
Training Step 1: Prepare Data . . . . . . . . . .
Training Step 2: Run GIZA++ . . . . . . . . . .
5.5.1 Training on really large corpora . . . .
5.5.2 Training in parallel . . . . . . . . . . . .
Training Step 3: Align Words . . . . . . . . . .
Training Step 4: Get Lexical Translation Table
Training Step 5: Extract Phrases . . . . . . . .
Training Step 6: Score Phrases . . . . . . . . .
Training Step 7: Build reordering model . . . .
Training Step 8: Build generation model . . . .
Training Step 9: Create Configuration File . . .
Building a Language Model . . . . . . . . . . .
5.13.1 Language Models in Moses . . . . . . .
5.13.2 Building a LM with the SRILM Toolkit
5.13.3 On the IRSTLM Toolkit . . . . . . . . .
5.13.4 RandLM . . . . . . . . . . . . . . . . . .
5.13.5 KenLM . . . . . . . . . . . . . . . . . .
Tuning . . . . . . . . . . . . . . . . . . . . . . .
5.14.1 Overview . . . . . . . . . . . . . . . . .
5.14.2 Batch tuning algorithms . . . . . . . . .
5.14.3 Online tuning algorithms . . . . . . . .
5.14.4 Tuning in Practice . . . . . . . . . . . .
7
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Background
6.1 Background . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Model . . . . . . . . . . . . . . . . . . . . .
6.1.2 Word Alignment . . . . . . . . . . . . . . .
6.1.3 Methods for Learning Phrase Translations
6.1.4 Och and Ney . . . . . . . . . . . . . . . . .
6.2 Decoder . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Translation Options . . . . . . . . . . . . .
6.2.2 Core Algorithm . . . . . . . . . . . . . . .
6.2.3 Recombining Hypotheses . . . . . . . . . .
6.2.4 Beam Search . . . . . . . . . . . . . . . . .
6.2.5 Future Cost Estimation . . . . . . . . . . .
6.2.6 N-Best Lists Generation . . . . . . . . . . .
6.3 Factored Translation Models . . . . . . . . . . . .
6.3.1 Motivating Example: Morphology . . . . .
6.3.2 Decomposition of Factored Translation . .
6.3.3 Statistical Model . . . . . . . . . . . . . . .
6.4 Confusion Networks Decoding . . . . . . . . . . .
6.4.1 Confusion Networks . . . . . . . . . . . .
6.4.2 Representation of Confusion Network . .
6.5 Word Lattices . . . . . . . . . . . . . . . . . . . . .
6.5.1 How to represent lattice inputs . . . . . . .
6.5.2 Configuring moses to translate lattices . .
6.5.3 Verifying PLF files with checkplf . . . . .
6.5.4 Citation . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
217
219
220
220
221
224
224
225
228
229
230
230
230
231
232
237
241
244
244
245
246
246
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
249
249
250
251
252
252
255
255
256
257
257
259
260
262
262
263
264
267
267
268
269
269
270
270
271
8
CONTENTS
6.6
7
Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Code Guide
7.1 Code Guide . . . . . . . . . . . . . . . . . . .
7.1.1 Github, branching, and merging . . .
7.1.2 The code . . . . . . . . . . . . . . . .
7.1.3 Quick Start . . . . . . . . . . . . . . .
7.1.4 Detailed Guides . . . . . . . . . . . .
7.2 Coding Style . . . . . . . . . . . . . . . . . .
7.2.1 Formatting . . . . . . . . . . . . . . .
7.2.2 Comments . . . . . . . . . . . . . . .
7.2.3 Data types and methods . . . . . . .
7.2.4 Source Control Etiquette . . . . . . .
7.3 Factors, Words, Phrases . . . . . . . . . . . .
7.3.1 Factors . . . . . . . . . . . . . . . . . .
7.3.2 Words . . . . . . . . . . . . . . . . . .
7.3.3 Factor Types . . . . . . . . . . . . . .
7.3.4 Phrases . . . . . . . . . . . . . . . . .
7.4 Tree-Based Model Decoding . . . . . . . . .
7.4.1 Looping over the Spans . . . . . . . .
7.4.2 Looking up Applicable Rules . . . . .
7.4.3 Applying the Rules: Cube Pruning .
7.4.4 Hypotheses and Pruning . . . . . . .
7.5 Multi-Threading . . . . . . . . . . . . . . . .
7.5.1 Tasks . . . . . . . . . . . . . . . . . . .
7.5.2 ThreadPool . . . . . . . . . . . . . . .
7.5.3 OutputCollector . . . . . . . . . . . .
7.5.4 Not Deleting Threads after Execution
7.5.5 Limit the Size of the Thread Queue .
7.5.6 Example . . . . . . . . . . . . . . . . .
7.6 Adding Feature Functions . . . . . . . . . . .
7.6.1 Video . . . . . . . . . . . . . . . . . .
7.6.2 Feature Function . . . . . . . . . . . .
7.6.3 Stateless Feature Function . . . . . .
7.6.4 Stateful Feature Function . . . . . . .
7.6.5 Place-holder features . . . . . . . . .
7.6.6 moses.ini . . . . . . . . . . . . . . . .
7.6.7 Examples . . . . . . . . . . . . . . . .
7.7 Adding Sparse Feature Functions . . . . . .
7.7.1 Implementation . . . . . . . . . . . .
7.7.2 Weights . . . . . . . . . . . . . . . . .
7.8 Regression Testing . . . . . . . . . . . . . . .
7.8.1 Goals . . . . . . . . . . . . . . . . . .
7.8.2 Test suite . . . . . . . . . . . . . . . .
7.8.3 Running the test suite . . . . . . . . .
7.8.4 Running an individual test . . . . . .
7.8.5 How it works . . . . . . . . . . . . . .
7.8.6 Writing regression tests . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
273
273
273
276
277
277
277
277
278
279
280
280
280
280
281
281
281
282
283
285
287
288
289
290
291
291
292
292
294
294
295
297
298
299
299
300
303
304
304
305
305
305
306
306
306
307
CONTENTS
8
Reference
8.1 Frequently Asked Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.1 My system is taking a really long time to translate a sentence. What can
I do to speed it up ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.2 The system runs out of memory during decoding. . . . . . . . . . . . . .
8.1.3 I would like to point out a bug / contribute code. . . . . . . . . . . . . . .
8.1.4 How can I get an updated version of Moses ? . . . . . . . . . . . . . . . .
8.1.5 What changed in the latest release of Moses? . . . . . . . . . . . . . . . . .
8.1.6 I am an undergrad/masters student looking for a project in SMT. What
should I do? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.7 What do the 5 numbers in the phrase table mean? . . . . . . . . . . . . . .
8.1.8 What OS does Moses run on? . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.9 Can I use Moses on Windows ? . . . . . . . . . . . . . . . . . . . . . . . .
8.1.10 Do I need a computer cluster to run experiments? . . . . . . . . . . . . . .
8.1.11 I have compiled Moses, but it segfaults when running. . . . . . . . . . . .
8.1.12 How do I add a new feature function to the decoder? . . . . . . . . . . . .
8.1.13 Compiling with SRILM or IRSTLM produces errors. . . . . . . . . . . . .
8.1.14 I am trying to use Moses to create a web page to do translation. . . . . . .
8.1.15 How can a create a system that translate both ways, ie. X-to-Y as well as
Y-to-X ? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.16 PhraseScore dies with signal 11 - why? . . . . . . . . . . . . . . . . . . . .
8.1.17 Does Moses do Hierarchical decoding, like Hiero etc? . . . . . . . . . . .
8.1.18 Can I use Moses in proprietary software ? . . . . . . . . . . . . . . . . . .
8.1.19 GIZA++ crashes with error "parameter ’coocurrencefile’ does not exist." .
8.1.20 Running regenerate-makefiles.sh gives me lots of errors about *GREP
and *SED macros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.21 Running training I got the following error "*** buffer overflow detected
***: ../giza-pp/GIZA++-v2/GIZA++ terminated" . . . . . . . . . . . . . .
8.1.22 I retrained my model and got different BLEU scores. Why? . . . . . . . .
8.1.23 I specified ranges for mert weights, but it returned weights which are
outwith those ranges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.24 Who do I ask if my question has not been answered by this FAQ? . . . .
8.2 Reference: All Decoder Parameters . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3 Reference: All Training Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.1 Basic Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.2 Factored Translation Model Settings . . . . . . . . . . . . . . . . . . . . . .
8.3.3 Lexicalized Reordering Model . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.4 Partial Training . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.5 File Locations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.6 Alignment Heuristic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.7 Maximum Phrase Length . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.8 GIZA++ Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.3.9 Dealing with large training corpora . . . . . . . . . . . . . . . . . . . . . .
8.4 Glossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
309
309
309
309
309
310
310
310
310
310
311
311
311
311
311
312
312
312
313
313
313
314
314
314
314
314
315
316
317
319
319
319
320
321
321
321
322
322
10
CONTENTS
1
Introduction
1.1
Welcome to Moses!
Moses is a statistical machine translation system that allows you to automatically train translation models for any language pair. All you need is a collection of translated texts (parallel
corpus). Once you have a trained model, an efficient search algorithm quickly finds the highest
probability translation among the exponential number of choices.
1.2
Overview
1.2.1
Technology
Moses is an implementation of the statistical (or data-driven) approach to machine translation
(MT). This is the dominant approach in the field at the moment, and is employed by the online translation systems deployed by the likes of Google and Microsoft. In statistical machine
translation (SMT), translation systems are trained on large quantities of parallel data (from
which the systems learn how to translate small segments), as well as even larger quantities of
monolingual data (from which the systems learn what the target language should look like).
Parallel data is a collection of sentences in two different languages, which is sentence-aligned,
in that each sentence in one language is matched with its corresponding translated sentence in
the other language. It is also known as a bitext.
The training process in Moses takes in the parallel data and uses coocurrences of words and
segments (known as phrases) to infer translation correspondences between the two languages
of interest. In phrase-based machine translation, these correspondences are simply between
continuous sequences of words, whereas in hierarchical phrase-based machine translation
or syntax-based translation, more structure is added to the correspondences. For instance a
hierarchical MT system could learn that the German hat X gegessen corresponds to the English
ate X, where the Xs are replaced by any German-English word pair. The extra structure used in
these types of systems may or may not be derived from a linguistic analysis of the parallel data.
Moses also implements an extension of phrase-based machine translation know as factored
translation which enables extra linguistic information to be added to a phrase-based systems.
11
12
1. Introduction
phrase-based MT (Section 3.1), syntactic MT (Section 3.3) or factored MT (Section 3.2).
Whichever type of machine translation model you use, the key to creating a good system is lots
of good quality data. There are many free sources of parallel data1 which you can use to train
sample systems, but (in general) the closer the data you use is to the type of data you want to
translate, the better the results will be. This is one of the advantages to using on open-source
tool like Moses, if you have your own data then you can tailor the system to your needs and
potentially get better performance than a general-purpose translation system. Moses needs
sentence-aligned data for its training process, but if data is aligned at the document level, it can
often be converted to sentence-aligned data using a tool like hunalign2
1.2.2
Components
The two main components in Moses are the training pipeline and the decoder. There are also
a variety of contributed tools and utilities. The training pipeline is really a collection of tools
(mainly written in perl, with some in C++) which take the raw data (parallel and monolingual)
and turn it into a machine translation model. The decoder is a single C++ application which,
given a trained machine translation model and a source sentence, will translate the source
sentence into the target language.
The Training Pipeline
There are various stages involved in producing a translation system from training data, which
are described in more detail in the training documentation (Section 5.1) and in the baseline
system guide (Section 2.3). These are implemented as a pipeline, which can be controlled by
the Moses experiment management system (Section 3.5), and Moses in general makes it easy
to insert different types of external tools into the training pipeline.
The data typically needs to be prepared before it is used in training, tokenising the text and converting tokens to a standard case. Heuristics are used to remove sentence pairs which look to
be misaligned, and long sentences are removed. The parallel sentences are then word-aligned,
typically using GIZA++3 , which implements a set of statistical models developed at IBM in the
80s. These word alignments are used to extract phrase-phrase translations, or hierarchical rules
as required, and corpus-wide statistics on these rules are used to estimate probabilities.
An important part of the translation system is the language model, a statistical model built
using monolingual data in the target language and used by the decoder to try to ensure the
fluency of the output. Moses relies on external tools (Section 5.13) for language model building.
The final step in the creation of the machine translation system is tuning (Section 5.14), where
the different statistical models are weighted against each other to produce the best possible
translations. Moses contains implementations of the most popular tuning algorithms.
1
http://mokk.bme.hu/resources/hunalign/
3
2
1.2.
Overview
13
The Decoder
The job of the Moses decoder is to find the highest scoring sentence in the target language
(according to the translation model) corresponding to a given source sentence. It is also possible for the decoder to output a ranked list of the translation candidates, and also to supply
various types of information about how it came to its decision (for instance the phrase-phrase
correspondences that it used).
The decoder is written in a modular fashion and allows the user to vary the decoding process
in various ways, such as:
• Input: This can be a plain sentence, or it can be annotated with xml-like elements to guide
the translation process, or it can be a more complex structure like a lattice or confusion
network (say, from the output of speech recognition)
• Translation model: This can use phrase-phrase rules, or hierarchical (perhaps syntactic)
rules. It can be compiled into a binarised form for faster loading. It can be supplemented
with features to add extra information to the translation process, for instance features
which indicate the sources of the phrase pairs in order to weight their reliability.
• Decoding algorithm: Decoding is a huge search problem, generally too big for exact
search, and Moses implements several different strategies for this search, such as stackbased, cube-pruning, chart parsing etc.
• Language model: Moses supports several different language model toolkits (SRILM,
KenLM, IRSTLM, RandLM) each of which has there own strengths and weaknesses, and
adding a new LM toolkit is straightforward.
The Moses decoder also supports multi-threaded decoding (since translation is embarassingly
parallelisable4 ), and also has scripts to enable multi-process decoding if you have access to a
cluster.
Contributed Tools
There are many contributed tools in Moses which supply additional functionality over and
above the standard training and decoding pipelines. These include:
• Moses server: which provides an xml-rpc interface to the decoder
• Web translation: A set of scripts to enable Moses to be used to translate web pages
• Analysis tools: Scripts to enable the analysis and visualisation of Moses output, in comparison with a reference.
There are also tools to evaluate translations, alternative phrase scoring methods, an implementation of a technique for weighting phrase tables, a tool to reduce the size of the phrase table,
and other contributed tools.
1.2.3
Development
Moses is an open-source project, licensed under the LGPL5 , which incorporates contributions
from many sources. There is no formal management structure in Moses, so if you want to
4
5
http://en.wikipedia.org/wiki/Embarrassingly_parallel
http://www.gnu.org/copyleft/lesser.html
14
1. Introduction
contribute then just mail support6 and take it from there. There is a list (Section 1.3) of possible
projects on this website, but any new MT techiques are fair game for inclusion into Moses.
repository, preferring the approach of removing/fixing bad commits, rather than vetting commits as they come in. This means that trunk occasionally breaks, but given the active Moses
user community, it doesn’t stay broken for long. The nightly builds and tests of trunk are reported on the cruise control7 web page, but if you want a more stable version then look for one
of the releases (Section 2.4).
1.2.4
Moses in Use
The liberal licensing policy in Moses, together with its wide coverage of current SMT technology and complete tool chain, make it probably the most widely used open-source SMT system.
It is used in teaching, research, and, increasingly, in commercial settings.
Commercial use of Moses is promoted and tracked by TAUS8 . The most common current use
for SMT in commercial settings is post-editing where machine translation is used as a firstpass, with the results then being edited by human translators. This can often reduce the time
(and hence total cost) of translation. There is also work on using SMT in computer-aided
translation, which is the research topic of two current EU projects, Casmacat9 and MateCat10 .
1.2.5
2005
2006
2006
2007
2009
2010
2011
2012
History
Hieu Hoang (then student of Philipp Koehn) starts Moses as successor to Pharoah
Moses is the subject of the JHU workshop, first check-in to public repository
Start of Euromatrix, EU project which helps fund Moses development
First machine translation marathon held in Edinburgh
Moses receives support from EuromatrixPlus, also EU-funded
Moses now supports hierarchical and syntax-based models, using chart decoding
Moses moves from sourceforge to github, after over 4000 sourceforge check-ins
EU-funded MosesCore launched to support continued development of Moses
1.3
Get Involved
1.3.1
Mailing List
The main forum for communication on Moses is the Moses support mailing list11 .
6
http://www.statmt.org/moses/?n=Moses.MailingLists
http://www.statmt.org/moses/cruise/
8
http://www.translationautomation.com/user-cases/machines-takes-center-stage.html
9
http://www.casmacat.eu/
10
http://www.matecat.com/
11
http://www.statmt.org/moses/?n=Moses.MailingLists
7
1.3.
Get Involved
1.3.2
15
Suggestions
We’d like to hear what you want from Moses. We can’t promise to implement the suggestions,
but they can be used as input into research and student projects, as well as Marathon12 projects.
If you have a suggestion/wish for a new feature or improvement, then either report them
via the issue tracker13 , contact the mailing list or drop Barry or Hieu a line (addresses on the
mailing list page).
1.3.3
Development
Moses is an open source project that is at home in the academic research community. There are
several venues where this community gathers, such as:
• The main conferences in the field: ACL, EMNLP, MT Summit, etc.
• The annual ACL Workshop on Statistical Machine Translation14
• The annual Machine Translation Marathon15
Moses is being developed as a reference implementation of state-of-the-art methods in statistical machine translation. Extending this implementation may be the subject of undergraduate or
graduate theses, or class projects. Typically, developers extend functionality that they required
for their projects, or to explore novel methods. Let us know if you made an improvement, no
matter how minor. Also let us know if you found or fixed a bug.
1.3.4
Use
We are aware of many commercial deployments of Moses, for instance as described by TAUS16 .
Please let us know if you use Moses commercially. Do not hesitate to contact the core developers of Moses. They are willing to answer questions and may be even available for consulting
services.
1.3.5
Contribute
There are many ways you can contribute to Moses.
• To get started, build systems with your data and get familiar with how Moses works.
• Test out alternative settings for building a system. The shared tasks organized around
the ACL Workshop on Statistical Machine Translation17 are a good forum to publish such
results on standard data conditions.
(Section 7.1) to make it easier for others to understand the code.
• If you come across inefficient implementations (e.g., bad algorithms or code in Perl that
should be ported to C++), program more efficient implementations.
• If you have new ideas for features, tools, and functionality, add them.
• Help out with some of the projects listed below.
12
http://www.statmt.org/moses/?n=Moses.Marathons
https://github.com/moses-smt/mosesdecoder/issues
14
http://www.statmt.org/wmt14/
15
http://www.statmt.org/moses/?n=Moses.Marathons
16
http://www.translationautomation.com/user-cases/machines-takes-center-stage.html
17
http://www.statmt.org/wmt14/
13
16
1. Introduction
1.3.6
Projects
If you are looking for projects to improve Moses, please consider the following list:
Popular Projects
• OpenOffice/Microsoft Word, Excel or Access plugins (Hieu Hoang) Create wrappers for
the Moses decoder to translate within user apps.
• Firefox, Chrome, Internet Explorer plugins (Hieu Hoang) Create a plugin that calls the
Moses server to translate webpages.
• Moses on the OLPC (Hieu Hoang) Create a front-end for the decoder, and possible the
training pipeline, so that it can be run on the OLPC. Some preliminary work has been
done here18
• Rule-Based Numbers, Currency, Date Translation (Hieu Hoang) SMT is bad at translating numbers and dates. Write some simple rules to identify and translate these for the
language pairs of your choice. Integrate it into Moses and combine it with the placeholder
feature19 .
• Named Entity Translation (Hieu Hoang) Text with lots of names and trademarks etc are
difficult for SMT to translate. Integrate named entity recognition into Moses. Translate
them using the transliteration phrase-table, placeholder feature, or a secondary phrasetable.
• Interactive visualization for SCFG decoding (Hieu Hoang) Create a front-end to the hiero/syntax decoder that enables the user to re-translate a part of the sentence, change
parameters in the decoder, add or delete translation rules etc.
Models
• Out-of-Vocabulary (OOV) Word Handing: Currently there are two choices for OOVs
- pass them through or drop them. Often neither is appropriate and Moses lacks good
hooks to add new OOV strategies, and lacks alternative strategies. A new phrase-table
class should be created which process OOV. To create a new phrase-table type, make a
copy of moses/TranslationModel/SkeletonPT.*, rename the class and follow the example in the file to implement your own code
• Tokenization for your language: Tokenization is the only part of the basic SMT process
that is language-specific. You can help make translation for your language better. Make
a copy of the file scripts/share/nonbreaking_prefixes/nonbreaking_prefix.en and
replace it with non-breaking words in your language.
Training
• Incremental updating of translation and language model: When you add new sentences
to the training data, you don’t want to re-run the whole training pipeline (do you?). Abby
Levenberg has implemented incremental training20 for Moses but what it lacks is a nice
How-To guide.
18
http://wiki.laptop.org/go/Projects/Automatic_translation_software
20
19
1.3.
Get Involved
17
• Faster tuning by reuse: In tuning, you constantly re-decode the same set of sentences
and this can be very time-consuming. What if you could reuse part of the calculation
each time? This has been previously proposed as a marathon project21
• Use binary files to speed up phrase scoring: Phrase-extraction and scoring involves a lot
of processing of text files which is inefficient in both time and disk usage. Using binary
files and vocabulary ids has the potential to make training more efficient, although more
opaque.
• Lattice training: At the moment lattices can be used for decoding (Section 6.5), and also
for MERT22 but they can’t be used in training. It would be pretty cool if they could be
used for training, but this is far from trivial.
Chart Decoding
• Decoding algorithms for syntax-based models: Moses generally supports a large set
of grammar types. For some of these (for instance ones with source syntax, or a very
large set of non-terminals), the implemented CKY decoding algorithm is not optimal.
Implementing search algorithms for dedicated models, or just to explore alternatives,
would be of great interest.
• Cube pruning for factored models: Complex factored models with multiple translation
and generation steps push the limits of the current factored model implementation which
exhaustively computes all translations options up front. Using ideas from cube pruning
(sorting the most likely rules and partial translation options) may be the basis for more
efficient factored model decoding.
• Missing features for chart decoder: a number of features are missing for the chart decoder, such as: MBR decoding (should be simple) and lattice decodings. In general, reporting and analysis within experiment.perl could be improved.
• More efficient rule table for chart decoder: The in-memory rule table for the hierarchical
decoder loads very slowly and uses a lot of RAM. A optimized implementation that is
vastly more efficient on both fronts should be feasible.
• Only maintain total hypothesis weight in decoding: At the moment, each hypothesis
(partial translation) contains the full feature vector, but really all that is required is the
weighted score. The feature vectors could then be supplied lazily, if needed for n-best
lists, and decoding would be more efficient.
• Scope-0 grammar and phrase-table (Hieu Hoang). The most popular decoding algorithm
for syntax MT is the CKY+ algorithm. This is a parsing algorithm which is able to use decoding with an unnormalized, unpruned grammar. However, the disadvantage of using
such a general algorithm is its speed; Hopkins and Langmead (2010) showed that that a
sentence of length n can be parsed using a scope-k grammar in O(nk) chart update. For
an unpruned grammar with 2 non-terminals (the usual SMT setup), the scope is 3.
This project proposes to quantify the advantages and disadvantages of scope-0 grammar. A
scope-0 grammar lacks application ambiguity, therefore, decoding can be fast and memory
efficient. However, this must be offset against potential translation quality degradation due to
the lack of coverage.
It may be that the advantages of a scope-0 grammar can only be realized through specifically
developed algorithms, such as parsing algorithms or data structures. The phrase-table lookup
21
22
http://www.statmt.org/mtm12/index.php%3Fn=Projects.TargetHypergraphSerialization
18
1. Introduction
for a Scope-0 grammar can be significantly simplified, made faster, and applied to much large
span width.
This project will also aim to explore this potentially rich research area.
http://aclweb.org/anthology//W/W12/W12-3150.pdf http://www.sdl.com/Images/emnlp2009_tcm1026628.pdf
Phrase-based Models
• Faster training for the global lexicon model: Moses implements the global lexicon model
proposed by Mauser et al.23 , but training features for each target word using a maximum
entropy trainer is very slow (years of CPU time). More efficient training or accommodation of training of only frequent words would be useful.
• A better phrase table: The current binarised phrase table suffers from (i) far too many layers of indirection in the code making it hard to follow and inefficient (ii) a cache-locking
mechanism which creates excessive contention; and (iii) lack of extensibility meaning that
(e.g.) word alignments were added on by extensively duplicating code. A new phrase
table could make Moses faster and more extensible.
• Multi-threaded Decoding: Moses uses a simple "thread per sentence" model for multithreaded decoding. However this means that if you have a single sentence to decode,
then multi-threading will not get you the translation any faster. Is it possible to have a
finer-grained threading model that can use multiple threads on a single sentence? This
would call for a new approach to decoding.
• Soft Constraints on Reordering: Moses currently allows you to specify hard constraints24
on reordering, but it might be useful to have "soft" versions of these constraints. This
would mean that the translation would incur a trainable penalty for violating the constraints, implemented by adding a feature function.
• Merging the Phrase-Table and Lexicalized Reordering Table (Hieu Hoang) They contain
the same source and target phrases, but different probabilities, and how those probabilities are applied. Merging the 2 models would halve the number of lookups.
• Sparse Reordering Features: Implementation of Cherry’s Improved Reordering for PhraseBased Translation using Sparse Features (NAACL 2013)25 .
• Using Artificial Neural Networks as Memory to Store the Phrase Table (Hieu Hoang)
ANN can be used as associative memory to store information in a lossy method. [http://ieeexplore.ieee.org
It would be interesting to use them to how useful they are at store the phrase table. Further research can focus on how they can be used to store morphologically similar translations.
Improved Tools
• Python Interface: A Python interface to the decoder could enable easy experimentation
and incorporation into other tools. cdec has one26 and Moses has a python interface to
the on-disk phrase tables (implemented by Wilker Aziz) but it would be useful to be able
to call the decoder from python.
23
http://aclweb.org/anthology/D/D09/D09-1022.pdf
25
http://www.aclweb.org/anthology-new/N/N13/N13-1003.pdf
26
http://ufal.mff.cuni.cz/pbml/98/art-chahuneau-smith-dyer.pdf
24
1.3.
Get Involved
19
• Analysis of results: (Philipp Koehn) Assessing the impact of variations in the design of
a machine translation system by observing the fluctuations of the BLEU score may not
be sufficiently enlightening. Having more analysis of the types of errors a system makes
should be very useful.
Engineering Improvements
• Integration of sigfilter: The filtering algorithm of Johnson et al27 is available28 in Moses,
but it is not well integrated, has awkward external dependencies and so is seldom used.
At the moment the code is in the contrib directory. A useful project would be to refactor
this code to use the Moses libraries for suffix arrays, and to integrate it with the Moses
experiment management system (ems). The goal would be to enable the filtering to be
turned on with a simple switch in the ems config file.
• Boostification: Moses has allowed boost29 since Autumn 2011, but there are still many
areas of the code that could be improved by usage of the boost libraries, for instance using
shared pointers in collections.
• Unit-testing: The core of Moses is almost completely lacking in unit testing, although
some exist for MERT and KenLM (using boost test). Increasing test coverage is a priority
for 2012, and implementing unit tests is a good way of learning about the code ;-). Some
refactoring will be necessary in order to make Moses "unit-testable".
• Cruise control Moses has cruise control30 running on a server at the University of Edinburgh, however this only tests one platform (OpenSuse). If you have a different platform,
and care about keeping Moses stable on that platform, then you could set up a cruise control instance too. The code is all in the standard Moses distribution.
• Multi-pass decoding (Hieu Hoang) Moses scores all hypotheses when they are created.
In some situations, it would be better to have all the hypotheses, then score them with a
feature function. For example, coarse-to-fine decoding. The are some major issues that
has to be tackled with multi-pass decoding:
1. A losing hypothesis which have been recombined with the winning hypothesis may now
be the new winning hypothesis. The output search graph has to be reordered to reflect
this.
2. The feature functions in the 2nd pass produce state information. Recombined hypotheses
may no longer be recombined and have to be split.
3. It would be useful for feature functions scores to be able to be evaluated asynchronously.
That is, a function to calculate the score it called but the score is calculated later.
Documentation
• Maintenance The documentation always needs maintenance as new features are introduced and old ones are updated. Such a large body of documentation inevitably contains
mistakes and inconsistencies, so any help in fixing these would be most welcome. If you
want to work on the documentation, just introduce yourself on the mailing list.
27
http://aclweb.org/anthology/D/D07/D07-1103.pdf
29
http://www.boost.org
30
http://www.statmt.org/moses/cruise/
28
20
1. Introduction
• Help Messages Moses has a lot of executables, and often the help messages are quite
cryptic or missing. A help message in the code is more likely to be maintained than
separate documentation, and easier to locate when you’re trying to find the right options.
Fixing the help messages would be a useful contribution to making Moses easier to use.
2
Installation
2.1
Getting Started with Moses
This section will show you how to install and build Moses, and how to use Moses to translate
with some simple models. If you experience problems, then please check the support1 page. If
you do not want to build Moses from source, then there are packages2 available for Windows
and popular Linux distributions.
2.1.1
It’s best to download the source code via git from github3 .
2.1.2
Basic Setup
To compile Moses, you need the following installed on your machine:
g++
Boost
It’s best to install these using the software manager of your operating system.
Once these are installed, run bjam.
This is the exact commands I (Hieu) use to download and install Moses:
1
http://www.statmt.org/moses/?n=Moses.MailingLists
http://www.statmt.org/moses/?n=Moses.Packages
3
https://github.com/moses-smt/mosesdecoder
2
21
22
2. Installation
git clone https://github.com/moses-smt/mosesdecoder.git
cd mosesdecoder/
./bjam -j8
You see what options are available with bjam, run
./bjam --help
Manually installing Boost
Boost 1.48 has a serious bug which breaks Moses compilation. Unfornately, some Linux distributions (eg. Ubuntu 12.04) have broken versions of the Boost library. In these cases, you must
download and compile Boost yourself. The instructions to do this are in the Moses file:
BUILD-INSTRUCTIONS.txt
This is the exact commands I use to compile boost:
tar zxvf boost_1_55_0.tar.gz
cd boost_1_55_0/
./bootstrap.sh
./b2 -j8 --prefix=$PWD --libdir=$PWD/lib64 --layout=system link=static install || echo FAILURE
This create library file in the directory lib64, NOT in the system directory.
Once boost is installed, you can then compile Moses. However, you must tell Moses where
boost is with the --with-boost flag. This is the exact commands I use to compile Moses:
./bjam --with-boost=~/workspace/temp/boost_1_55_0 -j8
2.1.
2.1.3
Getting Started with Moses
23
Other software to install
Word Alignment
Moses requires a word alignment tool, such as giza++4 , mgiza5 , or Fast Align6 .
I (Hieu) use MGIZA because it is multi-threaded and give general good result, however, I’ve
also heard good things about Fast Align. You can find instructions to compile them here7 .
Language Model Creation
Moses include the KenLM language model creation program, lmplz8 .
However, you can also create language models with IRSTLM9 and SRILM10 .
I (Hieu) regularly use IRSTLM to create language models because it is
1. free LGPL
2. runs on multiple-cores
3. can prune the LM.
If you just want to use IRSTLM to create language models, you don’t have to compile it into
Moses. You just need to call the IRSTLM create LM scripts.
However, if you want to use IRSTLM to read the language model and assign scores to n-grams,
then IRSTLM needs to be linked with Moses.
Once IRSTLM is successfully compiled, use the --with-irstlm switch to compile Moses with
IRSTLM. This is the exact command I use:
./bjam --with-irstlm=/home/s0565741/workspace/temp/irstlm-5.80.03 -j8
Personally, I only use IRSTLM as a query tool in this way if the LM n-gram order is over 7. In
most situation, I use KenLM because KenLM is multi-threadable and faster.
4
http://sourceforge.net/projects/mgizapp/
6
7
http://www.statmt.org/moses/?n=Moses.ExternalTools#ntoc3
8
http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#ntoc18
9
http://sourceforge.net/projects/irstlm/
10
11
http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#ntoc4
5
24
2. Installation
2.1.4
Platforms
The primary development platform for Moses is Linux, and this is the recommended platform
since you will find it easier to get support for it. However Moses does work on other platforms:
2.1.5
Windows Installation
Moses can run on Windows under Cygwin. Installation is exactly the same as for Linux and
Mac. (Are you running it on Windows? If so, please give us feedback on how it works). Cygwin
is available as a 32-bit or a 64-bit application. Make sure you select the one that’s appropriate
for your machine. Remember 32-bit limits the size of the models Moses can use to a maximum
of 4GB.
Install the following packages via Cygwin:
boost
automake
libtool
gcc-g++
python
git
subversion
openssh
make
tcl
zlib0
zlib-devel
libbz2_devel
unzip
libexpat-devel
libcrypt-devel
Also, the nist-bleu script need a perl module called XML::Twig12 . Install this in cygwin with
the command
cpan
cpan XML::Twig
12
http://search.cpan.org/~mirod/XML-Twig-3.44/Twig.pm
2.1.
2.1.6
Getting Started with Moses
25
OSX Installation
Mac OSX is widely used by Moses developers and everything should run fine. Installation is
the same as for Linux.
Mac OSX out-of-the-box doesn’t have many programs that are critical to Moses, or different
version of standard GNU programs. For example, split, sort, zcat are incompatible BSDversions rather than GNU versions.
Therefore, Moses has been tested with Mac OSX with Mac Ports. Make sure you have this
2.1.7
Linux Installation
Debian
Install the following packages using the command
su
apt-get install [package name]
Packages:
git
subversion
make
libtool
gcc
g++
libboost-dev
tcl-dev
tk-dev
zlib1g-dev
libbz2-dev
python-dev
Ubuntu
Install the following packages using the command
sudo apt-get install [package name]
26
2. Installation
Packages:
g++
git
subversion
automake
libtool
zlib1g-dev
libboost-all-dev
libbz2-dev
liblzma-dev
python-dev
libtcmalloc-minimal0 (libtcmalloc-minimal4 on Ubuntu 14.04)
Fedora
Install the following packages using the command
su
yumm install [package name]
Packages:
git
subversion
automake
libtool
gcc-c++
zlib-devel
python-devel
bzip2-devel
2.1.8
Run Moses for the first time
2.1.
Getting Started with Moses
27
cd ~/mosesdecoder
tar xzf sample-models.tgz
cd sample-models
Run the decoder
cd ~/mosesdecoder/sample-models
~/mosesdecoder/bin/moses -f phrase-model/moses.ini < phrase-model/in > out
If everything worked out right, this should translate the sentence "das ist ein kleines haus" (in
the file in) as "this is a small house" (in the file out).
Note that the configuration file moses.ini in each directory is set to use the KenLM language
model toolkit by default. If you prefer to use IRSTLM13 , then edit the language model entry in moses.ini, replacing KENLM with IRSTLM. You will also have to compile with ./bjam
Moses also supports SRILM and RandLM language models. See here14 for more details.
2.1.9
Chart Decoder
The chart decoder is created as a separate executable:
~/mosesdecoder/bin/moses_chart
You can run the chart demos from the sample-models directory as follows
~/mosesdecoder/bin/moses_chart -f string-to-tree/moses.ini < string-to-tree/in > out.stt
~/mosesdecoder/bin/moses_chart -f tree-to-tree/moses.ini < tree-to-tree/in.xml > out.ttt
The expected result of the string-to-tree demo is
this is a small house
13
14
http://hlt.fbk.eu/en/irstlm
http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#ntoc1
28
2.1.10
2. Installation
Next Steps
Why not try to build a Baseline (Section 2.3) translation system with freely available data?
2.1.11
bjam options
This is a list of options to bjam. On a system with Boost installed in a standard path, none
should be required, but you may want additional functionality or control.
Optional packages
Language models In addition to KenLM and ORLM (which are always compiled):
--with-irstlm=/path/to/irstlm Path to IRSTLM installation
--with-randlm=/path/to/randlm Path to RandLM installation
--with-srilm=/path/to/srilm Path to SRILM installation.
If your SRILM install is non-standard, use these options:
--with-srilm-arch=arch Override the arch setting given by /path/to/srilm/sbin/machine-type
Other packages
--with-boost=/path/to/boost If Boost is in a non-standard location, specify it here. This directory is expected to contain include and lib or lib64.
--with-xmlrpc-c=/path/to/xmlrpc-c Specify a non-standard libxmlrpc-c installation path. Used
by Moses server.
--with-cmph=/path/to/cmph Path where CMPH is installed. Used by the compact phrase table
and compact lexical reordering table.
--with-regtest=/path/to/moses-regression-tests Run the regression tests using data from this
2.1.
Getting Started with Moses
29
Installation
--prefix=/path/to/prefix sets the install prefix [default is source root].
--bindir=/path/to/prefix/bin sets the bin directory [default is PREFIX/bin]
--libdir=/path/to/prefix/lib sets the lib directory [default is PREFIX/lib]
--includedir=/path/to/prefix/include installs headers. Does not install if missing. No argument defaults to PREFIX/include .
--install-scripts=/path/to/scripts copies scripts into a directory. Does not install if missing. No
argument defaults to PREFIX/scripts .
--git appends the git revision to the prefix directory.
Build Options
variant=release|debug|profile builds optimized (default), for debug, or for profiling
--static forces static linking (the default will fall back to shared)
debug-symbols=on|off include (default) or exclude debugging information also known as -g
--notrace compiles without TRACE macros
--enable-boost-pool uses Boost pools for the memory SCFG table
--enable-mpi switch on mpi (used for MIRA - one of the tuning algorithms)
--without-libsegfault does not link with libSegFault
--max-kenlm-order maximum ngram order that kenlm can process (default 6)
--max-factors maximum number of factors (default 4)
--unlabelled-source ignore source nonterminals (if you only use hierarchical or string-to-tree
models without source syntax)
Controlling the Build
-q quit on the first error
-a to build from scratch
-jNCPUS to compile in parallel --clean to clean 30 2.2 2. Installation Building with Eclipse There is a YouTube video showing you how to set up Moses with Eclipse. {\bf How to compile Moses with Eclipse\footnote{\sf https://www.youtube.com/watch?v=P43h827uLac&feature=youtu.be}} Moses comes with Eclipse project files for some of the C++ executables. Currently, there are project files for * * * * * moses-cmd (decoder) moses-chart-cmd (decoder) extract extract-rules extract-ghkm The Eclipse build is used primarily for development and debugging. It is not optmized and doesn’t have many of the options available in the bjam build. The advantage of using Eclipse is that it offers code-completion, and a GUI debugging environment. NB. The recent update of Mac OSX replaces g++ with clang. Eclipse doesn’t yet fully function with clang. Follow these instructions to build with Eclipse: * Use the version of Eclipse for C++. My (Hieu) current version is Eclipse Kepler. * Get the Moses source code git clone [email protected]:moses-smt/mosesdecoder.git cd mosesdecoder * Create softlinks to Boost, cmph, DALM, irstlm, randlm, srilm in the Moses root directory eg. ln -s ~/workspace/randlm * Create a new Eclipse workspace. The workspace {\em MUST} be in contrib/other-builds/ Eclipse should now be running. * Import all the Moses Eclipse project into the workspace. File >> Import >> Existing Projects into Workspace >> Select Root Directory >> ’other-builds’ >> Finished * Compile all projects. Project >> Build All Subsection last modified on September 17, 2014, at 07:06 PM 2.3. Baseline System 2.3 Baseline System 2.3.1 Overview 31 This guide assumes that you have successfully installed Moses (Section 2.1), and would like to see how to use parallel data to build a real phrase-based translation system. The process requires some familiarity with UNIX and, ideally, access to a Linux server. It can be run on a laptop, but could take about a day and requires at least 2G of RAM, and about 10G of free disk space (these requirements are just educated guesses, so if you have a different experience then please mail support15 ). If you want to save the effort of typing in all the commands on this page (and see how the pros manage their experiments), then skip straight to the experiment management system (Section 2.3.8) instructions below. But I’d recommend that you follow through the process manually, at least once, just to see how it all works. 2.3.2 Installation The minimum software requirements are: • Moses (obviously!) • GIZA++16 , for word-aligning your parallel corpus • IRSTLM17 , SRILM18 , or KenLM19 for language model estimation. IRSTLM and KenLM are LGPL licensed (like Moses) and therefore available for commercial use. The Moses tool-chain defaults to SRILM, but it requires an expensive licence for nonacademic use. For the purposes of this guide, I will assume that you’re going to install all the tools and data in your home directory (i.e. ~/), and that you’ve already downloaded and compiled Moses into ~/mosesdecoder. And you’re going to run Moses from there. Installing GIZA++ GIZA++ is hosted at Google Code20 , and a mirror of the original documentation can be found here21 . I recommend that you download the latest version via svn: svn checkout http://giza-pp.googlecode.com/svn/trunk/ giza-pp cd giza-pp make 15 http://www.statmt.org/moses/?n=Moses.MailingLists http://code.google.com/p/giza-pp/ 17 http://hlt.fbk.eu/en/irstlm 18 http://www.speech.sri.com/projects/srilm/download.html 19 http://kheafield.com/code/kenlm/estimation/ 20 https://code.google.com/p/giza-pp/ 21 http://www.statmt.org/moses/giza/GIZA++.html 16 32 2. Installation This should create the binaries ~/giza-pp/GIZA++-v2/GIZA++, ~/giza-pp/GIZA++-v2/snt2cooc.out and ~/giza-pp/mkcls-v2/mkcls. These need to be copied to somewhere that Moses can find them as follows cd ~/mosesdecoder mkdir tools cp ~/giza-pp/GIZA++-v2/GIZA++ ~/giza-pp/GIZA++-v2/snt2cooc.out \ ~/giza-pp/mkcls-v2/mkcls tools When you come to run the training, you need to tell the training script where GIZA++ was installed using the -external-bin-dir argument. train-model.perl -external-bin-dirHOME/mosesdecoder/tools
UPDATE - GIZA++ only compiles with gcc. If you’re using OSX Mavericks, you’ll have to
install gcc yourself. I (Hieu) recommend using MGIZA instead
Installing IRSTLM
IRSTLM is a language modelling toolkit from FBK, and is hosted on sourceforge22 . Again,
download a later version) the following commands should build and install IRSTLM:
tar zxvf irstlm-5.80.03.tgz
cd irstlm-5.80.03
./regenerate-makefiles.sh
./configure --prefix=HOME/irstlm make install You should now have several binaries and scripts in ~/irstlm/bin, in particular build-lm.sh 2.3.3 Corpus Preparation To train a translation system we need parallel data (text translated into two different languages) which is aligned at the sentence level. Luckily there’s plenty of this data freely available, and for this system I’m going to use a small (only 130,000 sentences!) data set released for the 2013 Workshop in Machine Translation. To get the data we want, we have to download the tarball and unpack it (into a corpus directory in our home directory) as follows 22 http://sourceforge.net/projects/irstlm/ 2.3. Baseline System 33 cd mkdir corpus cd corpus wget http://www.statmt.org/wmt13/training-parallel-nc-v8.tgz tar zxvf training-parallel-nc-v8.tgz If you look in the ~/corpus/training directory you’ll see that there’s data from news-commentary (news analysis from project syndicate) in various languages. We’re going to build a FrenchEnglish (fr-en) translation system using the news commentary data set, but feel free to use one of the other language pairs if you prefer. To prepare the data for training the translation system, we have to perform the following steps: • tokenisation: This means that spaces have to be inserted between (e.g.) words and punctuation. • truecasing: The initial words in each sentence are converted to their most probable casing. This helps reduce data sparsity. • cleaning: Long sentences and empty sentences are removed as they can cause problems with the training pipeline, and obviously mis-aligned sentences are removed. The tokenisation can be run as follows: ~/mosesdecoder/scripts/tokenizer/tokenizer.perl -l en \ < ~/corpus/training/news-commentary-v8.fr-en.en \ > ~/corpus/news-commentary-v8.fr-en.tok.en ~/mosesdecoder/scripts/tokenizer/tokenizer.perl -l fr \ < ~/corpus/training/news-commentary-v8.fr-en.fr \ > ~/corpus/news-commentary-v8.fr-en.tok.fr The truecaser first requires training, in order to extract some statistics about the text: ~/mosesdecoder/scripts/recaser/train-truecaser.perl \ --model ~/corpus/truecase-model.en --corpus \ ~/corpus/news-commentary-v8.fr-en.tok.en ~/mosesdecoder/scripts/recaser/train-truecaser.perl \ --model ~/corpus/truecase-model.fr --corpus \ ~/corpus/news-commentary-v8.fr-en.tok.fr Truecasing uses another script from the Moses distribution: 34 2. Installation ~/mosesdecoder/scripts/recaser/truecase.perl \ --model ~/corpus/truecase-model.en \ < ~/corpus/news-commentary-v8.fr-en.tok.en \ > ~/corpus/news-commentary-v8.fr-en.true.en ~/mosesdecoder/scripts/recaser/truecase.perl \ --model ~/corpus/truecase-model.fr \ < ~/corpus/news-commentary-v8.fr-en.tok.fr \ > ~/corpus/news-commentary-v8.fr-en.true.fr Finally we clean, limiting sentence length to 80: ~/mosesdecoder/scripts/training/clean-corpus-n.perl \ ~/corpus/news-commentary-v8.fr-en.true fr en \ ~/corpus/news-commentary-v8.fr-en.clean 1 80 Notice that the last command processes both sides at once. 2.3.4 Language Model Training The language model (LM) is used to ensure fluent output, so it is built with the target language (i.e English in this case). The IRSTLM documentation gives a full explanation of the commandline options, but the following will build an appropriate 3-gram language model, removing singletons, smoothing with improved Kneser-Ney, and adding sentence boundary symbols: mkdir ~/lm cd ~/lm ~/irstlm/bin/add-start-end.sh \ < ~/corpus/news-commentary-v8.fr-en.true.en \ > news-commentary-v8.fr-en.sb.en export IRSTLM=HOME/irstlm; ~/irstlm/bin/build-lm.sh \
-i news-commentary-v8.fr-en.sb.en
\
-t ./tmp -p -s improved-kneser-ney -o news-commentary-v8.fr-en.lm.en
~/irstlm/bin/compile-lm \
--text yes \
news-commentary-v8.fr-en.lm.en.gz \
news-commentary-v8.fr-en.arpa.en
You can directly create an IRSTLM binary LM (for faster loading in Moses) by replacing the last
command with the following:
2.3.
Baseline System
35
~/irstlm/bin/compile-lm news-commentary-v8.fr-en.lm.en.gz \
news-commentary-v8.fr-en.blm.en
You can transform an arpa LM (*.arpa.en file) into an IRSTLM binary LM as follows:
~/irstlm/bin/compile-lm \
news-commentary-v8.fr-en.arpa.en \
news-commentary-v8.fr-en.blm.en
or viceversa, you can transform an IRSTLM binary LM into an arpa LM as follows:
~/irstlm/bin/compile-lm \
--text yes \
news-commentary-v8.fr-en.blm.en \
news-commentary-v8.fr-en.arpa.en
~/mosesdecoder/bin/build_binary \
news-commentary-v8.fr-en.arpa.en \
news-commentary-v8.fr-en.blm.en
You can check the language model by querying it, e.g.
echo "is this an English sentence ?" | ~/mosesdecoder/bin/query news-commentary-v8.fr-en.blm.en Loading statistics: Name:query VmPeak:46788 kB VmRSS:30828 kB RSSMax:0 kB user:0 sys:0 CPU:0 real:0.012207 is=35 2 -2.6704 this=287 3 -0.889896 an=295 3 -2.25226 English=7286 1 -5.27842 sentence=4470 2 -2.69906 \ ?=65 1 -3.32728 </s>=21 2 -0.0308115 Total: -17.1481 OOV: After queries: Name:query VmPeak:46796 kB VmRSS:30828 kB RSSMax:0 kB \ \ \ 0 \ 36 2. Installation user:0 sys:0 CPU:0 real:0.0129395 Total time including destruction: Name:query VmPeak:46796 kB VmRSS:1532 kB user:0 sys:0 CPU:0 real:0.0166016 2.3.5 RSSMax:0 kB \ Training the Translation System Finally we come to the main event - training the translation model. To do this, we run wordalignment (using GIZA++), phrase extraction and scoring, create lexicalised reordering tables and create your Moses configuration file, all with a single command. I recommend that you create an appropriate directory as follows, and then run the training command, catching logs: mkdir ~/working cd ~/working nohup nice ~/mosesdecoder/scripts/training/train-model.perl -root-dir train -corpus ~/corpus/news-commentary-v8.fr-en.clean -f fr -e en -alignment grow-diag-final-and -reordering msd-bidirectional-fe -lm 0:3:HOME/lm/news-commentary-v8.fr-en.blm.en:8
-external-bin-dir ~/mosesdecoder/tools >& training.out &
\
\
\
\
If you have a multi-core machine it’s worth using the -cores argument to encourage as much
parallelisation as possible.
This took about 1.5 hours using 2 cores on a powerful laptop (Intel i7-2640M, 8GB RAM, SSD).
Once it’s finished there should be a moses.ini file in the directory ~/working/train/model.
You can use the model specified by this ini file to decode (i.e. translate), but there’s a couple
of problems with it. The first is that it’s very slow to load, but we can fix that by binarising the
phrase table and reordering table, i.e. compiling them into a format that can be load quickly.
The second problem is that the weights used by Moses to weight the different models against
each other are not optimised - if you look at the moses.ini file you’ll see that they’re set to
default values like 0.2, 0.3 etc. To find better weights we need to tune the translation system,
which leads us on to the next step...
2.3.6
Tuning
This is the slowest part of the process, so you might want to line up something to read whilst it’s
progressing. Tuning requires a small amount of parallel data, separate from the training data,
so again we’ll download some data kindly provided by WMT. Run the following commands
2.3.
Baseline System
37
cd ~/corpus
wget http://www.statmt.org/wmt12/dev.tgz
tar zxvf dev.tgz
We’re going to use news-test2008 for tuning, so we have to tokenise and truecase it first (don’t
forget to use the correct language if you’re not building a fr->en system)
cd ~/corpus
~/mosesdecoder/scripts/tokenizer/tokenizer.perl -l en \
< dev/news-test2008.en > news-test2008.tok.en
~/mosesdecoder/scripts/tokenizer/tokenizer.perl -l fr \
< dev/news-test2008.fr > news-test2008.tok.fr
~/mosesdecoder/scripts/recaser/truecase.perl --model truecase-model.en \
< news-test2008.tok.en > news-test2008.true.en
~/mosesdecoder/scripts/recaser/truecase.perl --model truecase-model.fr \
< news-test2008.tok.fr > news-test2008.true.fr
Now go back to the directory we used for training, and launch the tuning process:
cd ~/working
nohup nice ~/mosesdecoder/scripts/training/mert-moses.pl \
~/corpus/news-test2008.true.fr ~/corpus/news-test2008.true.en \
~/mosesdecoder/bin/moses train/model/moses.ini --mertdir ~/mosesdecoder/bin/ \
&> mert.out &
If you have several cores at your disposal, then it’ll be a lot faster to run Moses multi-threaded.
Add --decoder-flags="-threads 4" to the last line above in order to run the decoder with 4
The end result of tuning is an ini file with trained weights, which should be in ~/working/mertwork/moses.ini if you’ve used the same directory structure as me.
2.3.7
Testing
You can now run Moses with
~/mosesdecoder/bin/moses -f ~/working/mert-work/moses.ini
38
2. Installation
and type in your favourite French sentence to see the results. You’ll notice, though, that the
decoder takes at least a couple of minutes to start-up. In order to make it start quickly, we
can binarise the phrase-table and lexicalised reordering models. To do this, create a suitable
directory and binarise the models as follows:
mkdir ~/working/binarised-model
cd ~/working
~/mosesdecoder/bin/processPhraseTable \
-ttable 0 0 train/model/phrase-table.gz \
-nscores 5 -out binarised-model/phrase-table
~/mosesdecoder/bin/processLexicalTable \
-in train/model/reordering-table.wbe-msd-bidirectional-fe.gz \
-out binarised-model/reordering-table
Then make a copy of the ~/working/mert-work/moses.ini in the binarised-model directory
and change the phrase and reordering tables to point to the binarised versions, as follows:
1. Change PhraseDictionaryMemory to PhraseDictionaryBinary
2. Set the path of the PhraseDictionary feature to point to $HOME/working/binarisedmodel/phrase-table 3. Set the path of the LexicalReordering feature to point to$HOME/working/binarisedmodel/reordering-table
Loading and running a translation is pretty fast (for this I supplied the French sentence "faire
revenir les militants sur le terrain et convaincre que le vote est utile .") :
Defined parameters (per moses.ini or switch):
config: binarised-model/moses.ini
distortion-limit: 6
feature: UnknownWordPenalty WordPenalty PhraseDictionaryBinary
name=TranslationModel0 table-limit=20 num-features=5
input-factor=0 output-factor=0
LexicalReordering name=LexicalReordering0
num-features=6 type=wbe-msd-bidirectional-fe-allff
input-factor=0 output-factor=0
Distortion KENLM lazyken=0 name=LM0
input-factors: 0
mapping: 0 T 0
weight: LexicalReordering0= 0.119327 0.0221822 0.0359108
0.107369 0.0448086 0.100852 Distortion0= 0.0682159
LM0= 0.0794234 WordPenalty0= -0.0314219 TranslationModel0= 0.0477904
0.0621766 0.0931993 0.0394201 0.147903
\
\
\
\
\
\
\
\
\
\
2.3.
Baseline System
line=UnknownWordPenalty
FeatureFunction: UnknownWordPenalty0 start: 0 end: 0
line=WordPenalty
FeatureFunction: WordPenalty0 start: 1 end: 1
line=PhraseDictionaryBinary name=TranslationModel0 table-limit=20
input-factor=0 output-factor=0
FeatureFunction: TranslationModel0 start: 2 end: 6
line=LexicalReordering name=LexicalReordering0 num-features=6
type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0
FeatureFunction: LexicalReordering0 start: 7 end: 12
Initializing LexicalReordering..
line=Distortion
FeatureFunction: Distortion0 start: 13 end: 13
line=KENLM lazyken=0 name=LM0 factor=0 \
FeatureFunction: LM0 start: 14 end: 14
binary file loaded, default OFF_T: -1
IO from STDOUT/STDIN
Created input-output object : [0.000] seconds
Translating line 0 in thread id 140592965015296
Translating: faire revenir les militants sur le terrain et
convaincre que le vote est utile .
size of OFF_T 8
binary phrasefile loaded, default OFF_T: -1
binary file loaded, default OFF_T: -1
Line 0: Collecting options took 0.000 seconds
Line 0: Search took 1.000 seconds
bring activists on the ground and convince that the vote is useful .
BEST TRANSLATION: bring activists on the ground and convince that
the vote is useful . [111111111111111] [total=-8.127]
core=(0.000,-13.000,-10.222,-21.472,-4.648,-14.567,6.999,-2.895,0.000,
0.000,-3.230,0.000,0.000,0.000,-76.142)
Line 0: Translation took 1.000 seconds total
RSSMax:0 kB user:0.000 sys:0.000 CPU:0.000 real:1.031
39
\
\
\
\
\
\
\
\
\
The translation ("bring activists on the ground and convince that the vote is useful .")b is quite
rough, but understandable - bear in mind this is a very small data set for general domain translation. Also note that your results may differ slightly due to non-determinism in the tuning
process.
At this stage, your probably wondering how good the translation system is. To measure this,
we use another parallel data set (the test set) distinct from the ones we’ve used so far. Let’s pick
newstest2011, and so first we have to tokenise and truecase it as before
40
2. Installation
cd ~/corpus
~/mosesdecoder/scripts/tokenizer/tokenizer.perl -l en \
< dev/newstest2011.en > newstest2011.tok.en
~/mosesdecoder/scripts/tokenizer/tokenizer.perl -l fr \
< dev/newstest2011.fr > newstest2011.tok.fr
~/mosesdecoder/scripts/recaser/truecase.perl --model truecase-model.en \
< newstest2011.tok.en > newstest2011.true.en
~/mosesdecoder/scripts/recaser/truecase.perl --model truecase-model.fr \
< newstest2011.tok.fr > newstest2011.true.fr
The model that we’ve trained can then be filtered for this test set, meaning that we only retain
the entries needed translate the test set. This will make the translation a lot faster.
cd ~/working
~/mosesdecoder/scripts/training/filter-model-given-input.pl
\
filtered-newstest2011 mert-work/moses.ini ~/corpus/newstest2011.true.fr \
-Binarizer ~/mosesdecoder/bin/processPhraseTable
You can test the decoder by first translating the test set (takes a wee while) then running the
BLEU script on it:
nohup nice ~/mosesdecoder/bin/moses
-f ~/working/filtered-newstest2011/moses.ini
< ~/corpus/newstest2011.true.fr
> ~/working/newstest2011.translated.en
2> ~/working/newstest2011.out
~/mosesdecoder/scripts/generic/multi-bleu.perl
-lc ~/corpus/newstest2011.true.en
< ~/working/newstest2011.translated.en
\
\
\
\
\
\
This gives me a BLEU score of 23.5 (in comparison, the best result at WMT11 was 30.523 , although it should be cautioned that this uses NIST BLEU, which does its own tokenisation, so
there will be 1-2 points difference in the score anyway)
2.3.8
Experiment Management System (EMS)
If you’ve been through the effort of typing in all the commands, then by now you’re probably
wondering if there’s an easier way. If you’ve skipped straight down here without bothering
about the manual route then, well, you may have missed on a useful Moses "rite of passage".
23
http://matrix.statmt.org/matrix/systems_list/1669
2.4.
Releases
41
The easier way is, of course, to use the EMS (Section 3.5). To use EMS, you’ll have to install a
few dependencies, as detailed on the EMS page, and then you’ll need this config24 file. Make
a directory ~/working/experiments and place the config file in there. If you open it up, you’ll
see the home-dir variable defined at the top - then make the obvious change. If you set the
home directory, download the train, tune and test data and place it in the locations described
above, then this config file should work.
To run EMS from the experiments directory, you can use the command:
nohup nice ~/mosesdecoder/scripts/ems/experiment.perl -config config -exec &> log &
then sit back and wait for the BLEU score to appear in evaluation/report.1
2.4
Releases
2.4.1
Release 2.1.1 (3rd March, 2014)
This is a minor patch for a bug that prevent Moses from linking with tcmalloc when it is available on the compilation machine. Using tcmalloc can substantially speed up decoding, at the
cost of more memory usage.
• Get the code on github25
2.4.2
Release 2.1 (21th Jan, 2014)
This is the current stable release.
•
•
•
•
24
Get the code on github26
Virtual Machines models29
https://github.com/moses-smt/mosesdecoder/tree/RELEASE-2.1
26
https://github.com/moses-smt/mosesdecoder/tree/RELEASE-2.1
27
http://www.statmt.org/moses/RELEASE-2.1/binaries/
28
http://www.statmt.org/moses/RELEASE-2.1/models/
29
http://www.statmt.org/moses/RELEASE-2.1/vm/
25
42
2. Installation
Overview
The broad aim of this release is to tackle more complicated issues to enable better expandability
and reliability.
Specifically, the decoder has been refactored to create a more modular framework to enable
easier incorporation of new feature functions into Moses. This has necessitate major changes
in many other parts of the toolkit, including training and tuning.
As well as the refactored code, this release also incorporate a host of new features donated
by other developers. Transliteration modules, better error handling, small and fast language
models, and placeholders are just some of the new features that spring to mind.
We have also continue to expand the testing regime to maintain the reliability of the toolkit,
while enable more developers to contribute to the project.
We distribute Moses as: 1. source code, 2. binaries for Windows (32 and 64 bit), Mac OSX
(Mavericks), and various flavours of Linux (32 and 64 bit). 3. pre-installed in a Linux virtual
machine, using the open source VirtualBox application. 4. Amazon cloud server image.
2.4.3
Release 1.0 (28th Jan, 2013)
• Get the code on github30
Overview
The Moses community has grown tremendously over the last few years. From the beginning
as a purely research-driven project, we are now a diverse community of academic and business
users, ranging in experience from hardened developers to new users.
Therefore, the first priority of this release has been to concentrate on resolving long-standing,
but straightforward, issues to make the toolkit easier to use and more efficient. The provision of
full-time development team devoted to the maintenance and enhancement of the Moses toolkit
has allowed has to tackle many useful engineering problems.
A second priority was to put in place a multi-tiered testing regime to enable more developers
to contribute to the project, more quickly, while ensuring the reliability of the toolkit. However,
we have not stopped adding new features to the toolkit; the next section lists a number of major
features added in the last 9 months.
New Features
The following is a list of the major new features in the Moses toolkit since May 2012, in roughly
chronological order.
30
https://github.com/moses-smt/mosesdecoder/tree/RELEASE-1.0
http://www.statmt.org/moses/RELEASE-1.0/binaries/
32
http://www.statmt.org/moses/RELEASE-1.0/models/
31
2.4.
Releases
43
Parallel Training by Hieu Hoang and Rohit Gupta. The training process has been improved
and can take advantage of multi-core machines. Parallelization was achieved by partitioning
the input data, then running the translation rule extraction processes in parallel before merging
the data. The following is the timing for the extract process on different number of cores:
Cores
Time taken (mins)
One
48:55
Two
33:13
Three
27:52
Four
25:35
The training processes have also been redesigned to decrease disk access, and to use less disk
space. This is important for parallel processing as disk IO often becomes the limiting factor
with a large number of simultaneous disk access. It is also important when training syntactically inspired models or using large amounts of training data, which can result in very large
translation models.
IRST LM training integration by Hieu Hoang and Philipp Koehn The IRST toolkit for
training language models have been integrated into the Experiment Management System. The
SRILM software previously carried out this functionality. Substituting IRST for SRI means that
the entire training pipeline can be run using only free, open-source software. Not only is the
IRST toolkit unencumbered by a proprietary license, it is also parallelizable and capable of
training with a larger amount of data than was otherwise possible with SRI.
Distributed Language Model by Oliver Wilson. Language models can be distributed across
many machines, allowing more data to be used at the cost of a performance overhead. This is
still experimental code.
Incremental Search Algorithm by Kenneth Heafield. A replacement for the cube pruning
algorithm in CKY++ decoding, used in hierarchical and syntax models. It offers better tradeoff
between decoding speed and translation quality.
Compressed Phrase-Table and Reordering-Tables by Marcin Junczys-Dowmunt. A phrasetable and lexicalized reordering-table implementation which is both small and fast. More details33 .
Sparse features by Eva Hasler, Barry Haddow, Philipp Koehn A framework to allow a large
number of sparse features in the decoder. A number of sparse feature functions described in
the literature have been reproduced in Moses. Currently, the available sparse feature functions
are:
1.
2.
3.
4.
5.
6.
33
TargetBigramFeature
TargetNgramFeature
SourceWordDeletionFeature
SparsePhraseDictionaryFeature
GlobalLexicalModelUnlimited
PhraseBoundaryState
http://www.staff.amu.edu.pl/~junczys/images/7/7b/Mjd2012tsd1.pdf
44
2. Installation
7. PhraseLengthFeature
8. PhrasePairFeature
9. TargetWordInsertionFeature
Suffix array for hierarchical models by Hieu Hoang The training of syntactically-inspired
hierarchical models requires a large amount of time and resource. An alternative to training a
translation is to only extract the required translation rules for each input sentence.
We have integrated Adam Lopez’s suffix array implementation into Moses. This is a wellknown and mature implementation, which is hosted and maintained by the cdec community.
Batched MIRA by Colin Cherry. A replacement for MERT, especially suited for tuning a
large number of sparse features. (Cherry and Foster, NAACL 201234 ).
LR score by Lexi Birch and others. The BLEU score commonly used in MT is insensitive
to reordering errors. We have integrated another metric , LR score, described in (Birch and
Osborne, 201135 ) which better accounts for reordering, in the Moses toolkit.
Convergence of Translation Memory and Statistical Machine Translation by Philipp Koehn
and Hieu Hoang An alternative extract algorithm, (Koehn, Senellart, 2010 AMTA36 ), which
is inspired by the use of translation memories has been integrated into the Moses toolkit.
Word Alignment Information is turned on by default by Hieu Hoang and Barry Haddow
The word alignment produced by GIZA++/mgiza is carried by the phrase-table and made
available to the decoder. This information is required by some feature functions. The use of
these word alignment is now optimized for memory and speed, and enabled by default.
Modified Moore-Lewis filtering by Barry Haddow and Philipp Koehn Reimplementation
of domain adaptation of parallel corpus described by Axelrod et al. (EMNLP 2011)37 .
Lots and lots of cleanups and bug fixes By Ales Tamchyna, Wilker Aziz, Mark Fishel, Tetsuo Kiso, Rico Sennrich, Lane Schwartz, Hiroshi Umemoto, Phil Williams, Tom Hoar, Arianna
Bisazza, Jacob Dlougach, Jonathon Clark, Nadi Tomeh, Karel Bilek, Christian Buck, Oliver Wilson, Alex Fraser, Christophe Servan, Matous Machecek, Christian Federmann, Graham Neubig.
34
http://aclweb.org/anthology/P/P11/P11-1103.pdf
36
http://www.mt-archive.info/AMTA-2010-Koehn.pdf
37
http://aclweb.org/anthology/D/D11/D11-1033.pdf
35
2.4.
Releases
45
Building and Installing
The structure and installation of the Moses toolkit has been simplified to make compilation and
installation easier. The training and decoding process can be run from the directory in which
and executed immediately. Previously, the installation needed to be configured specifically for
the user’s machine.
A new build system has been implemented to build the Moses toolkit. This uses the boost
library’s build framework. The new system offers several advantages over the previous build
system.
Firstly, the source code for the new build system is included in the Moses repository which
is then bootstrapped the first time Moses is compiled. It does not rely on the the cmake, automake, make, and libtool applications. These have issues with cross-platform compatibility
and running on older operating systems.
Secondly, the new build system integrates the running of the unit tests and regression tests
with compilation.
Thirdly, the new system is significantly more powerful, allowing us to support a number of
new build features such as static and debug compilation, linking to external libraries such as
mpi and tmalloc, and other non-standard builds.
Testing
The MosesCore team has implemented several layers of testing to ensure the reliability of the
toolkit. We describe each below.
Unit Testing Unit testing tests each function or class method in isolation. Moses uses the
unit testing framework available from the Boost library to implement unit testing.
The source code for the unit tests are integrated into the Moses source. The tests are executed
every time the Moses source is compiled.
The unit testing framework has recently been implemented. There are currently 20 unit tests
for various features in mert, mira, phrase extraction, and decoding.
Regression Testing The regression tests ensure that changes to source code do not have
unknown consequences to existing functionality. The regression tests are typically applied to a
larger body of work than unit tests. They are designed to test specific functionality rather than
a specific function. Therefore, regression tests are applied to the actual Moses programs, rather
than tested in isolation.
The regression test framework forms the core of testing within the Moses toolkit. However, it
was created many years ago at the beginning of the Moses project and was only designed to
test the decoder. During the past 6 months, the scope of the regression test framework has been
expanded to test any part of the Moses toolkit, in addition to testing the decoder. The test are
grouped into the following types:
46
2. Installation
1.
2.
3.
4.
5.
6.
Phrase-based decoder
Hierarchical/Syntax decoder
Mert
Rule Extract
Phrase-table scoring
Miscellaneous, including domain adaptation features, binarizing phrase tables, parallel
rule extract, and so forth.
The number of tests has increased from 46 in May 2012 to 73 currently.
We have also overhauled the regression test to make it easier to add new tests. Previously, the
data for the regression tests could only be updated by developers who had access to the web
server at Edinburgh University. This has now been changed so that the data now resides in a
versioned repository on github.com38 .
This can be accessed and changed by any Moses developer, and is subject to the same checks
and controls as the rest of the Moses source code.
Every Moses developer is obliged to ensure the regression test are successfully executed before
they commit their changes to the master repository.
Cruise Control This is a daily task run on a server at the University of Edinburgh which
compiles the Moses source code and executes the unit tests and regressions tests. Additionally, it also runs a small training pipeline to completion. The results of this testing is publicly
available online39 .
This provides an independent check that all unit tests and regression tests passed, and that the
entirety of the SMT pipeline is working. Therefore, it tests not only the Moses toolkit, but also
external tools such as GIZA++ that are essential to Moses and the wider SMT community.
All failures are investigated by the MosesCore team and any remedial action is taken. This is
done to enforce the testing regime and maintain reliability.
The cruise control is a subproject of Moses initiated by Ales Tamchyna with contribution by
Operating-System Compatibility
The Moses toolkit has always strived to be compatible on multiple platforms, particularly on
the most popular operating systems used by researchers and commercial users.
Before each release, we make sure that Moses compiles and the unit tests and regression test
successfully runs on various operating systems.
Moses, GIZA++ mgiza, and IRSTLM was compiled for
1.
2.
3.
4.
38
39
Linux 32-bit
Linux 64-bit
Cygwin
Mac OSX 10.7 64-bit
https://github.com/moses-smt/moses-regression-tests
http://www.statmt.org/moses/cruise/
2.4.
Releases
47
Effort was made to make the executables runnable on as many platforms as possible. Therefore,
they were statically linked when possible. Moses was then tested on the following platforms:
1.
2.
3.
4.
5.
6.
Windows 7 (32-bit) with Cygwin 6.1
Mac OSX 10.7 with MacPorts
Ubuntu 12.10, 32 and 64-bit
Debian 6.0, 32 and 64-bit
Fedora 17, 32 and 64-bit
openSUSE 12.2, 32 and 64-bit
compile their own version.
GIZA++, mgiza, and IRSTLM are also available for download as binaries to enable users to run
the entire SMT pipeline without having to download and compile their own software.
Issues:
1. IRSTLM was not linked statically. The 64-bit version fails to execute on Debian 6.0. All
2. Mac OSX does not support static linking. Therefore, it is not known if the executables
would work on any other platforms, other than the one on which it was tested.
3. mgiza compilation failed on Mac OSX with gcc v4.2. It could only be successfully compilednwith gcc v4.5, available via MacPorts.
End-to-End Testing Before each Moses release, a number of full scale experiments are run.
This is the final test to ensure that the Moses pipeline can run from beginning to end, uninterrupted, with "real-world" datasets. The translation quality, as measured by BLEU, is also noted,
to ensure that there is no decrease in performance due to any interaction between components
in the pipeline.
This testing takes approximately 2 weeks to run. The following datasets and experiments are
currently used for end-to-end testing:
•
•
•
•
•
Europarl en-es: phrase-based, hierarchical
Europarl en-es: phrase-based, hierarchical
Europarl cs-en: phrase-based, hierarchical
Europarl en-cs: phrase-based, hierarchical
Europarl de-en: phrase-based, hierarchical, factored German POS, factored German+English
POS
• Europarl en-de: phrase-based, hierarchical, factored German POS, factored German+English
POS
• Europarl fr-en: phrase-based, hierarchical, recased (as opposed to truecased), factored
English POS
• Europarl en-fr: phrase-based, hierarchical, recased (as opposed to truecased), factored
English POS
40
http://www.statmt.org/moses/RELEASE-1.0/binaries/
48
2. Installation
Pre-Made Models The end-to-end tests produces a large number of tuned models. The models, as well as all configuration and data files, are made available for download41 . This is useful
as a template for users setting up their own experimental environment, or for those who just
want the models without running the experiments.
2.4.4
Release 0.91 (12th October, 2012)
The code is available in a branch on github42 .
This version was tested on 8 Europarl language pairs, phrase-based, hierarchical, and phrasebase factored models. All runs through without major intervention. Known issues:
1. Hierarchical models crashes on evaluation when threaded. Strangely, run OK during
tuning
2. EMS bugs when specifying multiple language models
3. Complex factored models not tested
4. Hierarchical models with factors does not work
2.4.5
Status 11th July, 2012
A roundup of the new features that have been implemented in the past year:
1. Lexi Birch’s LR score integrated into tuning. Finished coding: YES. Tested: NO. Documented: NO. Developer: Hieu, Lexi. First/Main user: Yvette Graham.
2. Asynchronous, batched LM requests for phrase-based models. Finished coding: YES.
Tested: UNKNOWN. Documented: YES. Developer: Oliver Wilson, Miles Osborne. First/Main
user: Miles Osborne.
3. Multithreaded tokenizer. Finished coding: YES. Tested: YES. Documented: NO. Developer: Pidong Wang.
4. KB Mira. Finished coding: YES. Tested: YES. Documented: YES. Developer: Colin
Cherry.
5. Training & decoding more resilient to non-printing characters and Moses’ reserved characters. Escaping the reserved characters and throwing away lines with non-printing
chars. Finished coding: YES. Tested: YES. Documented: NO. Developer: Philipp Koehn
and Tom Hoar.
6. Simpler installation. Finished coding: YES. Tested: YES. Documented: YES. Developer:
Hieu Hoang. First/Main user: Hieu Hoang.
7. Factors work with chart decoding. Finished coding: YES. Tested: NO. Documented: NO.
Developer: Hieu Hoang. First/Main user: Fabienne Braune.
8. Less IO and disk space needed during training. Everything written directly to gz files.
Finished coding: YES. Tested: YES. Documented: NO. Developer: Hieu. First/Main user:
Hieu.
9. Parallel training. Finished coding: YES. Tested: YES. Documented: YES. Developer:
Hieu. First/Main user: Hieu
10. Adam Lopez’s suffix array integrated into Moses’s training & decoding. Finished coding:
YES. Tested: NO. Documented: YES. Developer: Hieu.
41
42
http://www.statmt.org/moses/RELEASE-1.0/models/
https://github.com/moses-smt/mosesdecoder/tree/RELEASE-0.91
2.4.
Releases
49
11. Major MERT code cleanup. Finished coding: YES. Tested: NO. Documented: NO. Developer: Tetsuo Kiso.
12. Wrapper for Berkeley parser (german). Finished coding: YES. Tested: UNKNOWN. Documented: UNKNOWN. Developer: Philipp Koehn.
13. Option to use p(RHS_t|RHS_s,LHS) or p(LHS,RHS_t|RHS_s), as a grammar rule’s direct translation score. Finished coding: YES. Tested: UNKNOWN. Documented: UNKNOWN. Developer: Philip Williams. First/Main user: Philip Williams.
14. Optional PCFG scoring feature for target syntax models. Finished coding: YES. Tested:
UNKNOWN. Documented: UNKNOWN. Developer: Philip Williams. First/Main user:
Philip Williams.
15. Add -snt2cooc option to use mgiza’s reduced memory snt2cooc program. Finished coding: YES. Tested: YES. Documented: YES. Developer: Hieu Hoang.
16. queryOnDiskPt program. Finished coding: YES. Tested: YES. Documented: NO. Developer: Hieu Hoang. First/Main user: Daniel Schaut.
17. Output phrase segmentation to n-best when -report-segmentation is used. Finished
coding: YES. Tested: UNKNOWN. Developer: UNKNOWN. First/Main user: Jonathon
Clark.
18. CDER and WER metric in tuning. Finished coding: UNKNOWN. Tested: UNKNOWN.
Documented: UNKNOWN. Developer: Matous Machacek.
19. Lossy Distributed Hash Table Language Model. Finished coding: UNKNOWN. Tested:
UNKNOWN. Documented: UNKNOWN. Developer: Oliver Wilson.
20. Interpolated scorer for MERT. Finished coding: YES. Tested: UNKNOWN. Documented:
UNKNOWN. Developer: Matous Machacek.
21. IRST LM training integrated into Moses. Finished coding: YES. Tested: YES. Documented: YES. Developer: Hieu Hoang.
22. GlobalLexiconModel. Finished coding: UNKNOWN. Tested: UNKNOWN. Documented:
UNKNOWN. Developer: Jiri Marsik, Christian Buck and Philipp Koehn.
23. TM Combine (translation model combination). Finished coding: YES. Tested: YES. Documented: YES. Developer: Rico Sennrich.
24. Alternative to CKY+ for scope-3 grammar. Reimplementation of Hopkins and Langmead
(2010). Finished coding: YES. Tested: UNKNOWN. Documented: UNKNOWN. Developer: Philip Williams.
25. Sample Java client for Moses server. Finished coding: YES. Tested: NO. Documented:
NO. Developer: Marwen Azouzi. First/Main user: Mailing list users.
26. Support for mgiza, without having to install GIZA++ as well. Finished coding: YES.
Tested: YES. Documented: NO. Developer: Marwen Azouzi.
27. Interpolated language models. Finished coding: YES. Tested: YES. Documented: YES.
Developer: Philipp Koehn.
28. Duplicate removal in MERT. Finished coding: YES. Tested: YES. Documented: NO. Developer: Thomas Schoenemann.
29. Use bjam instead of automake to compile. Finished coding: YES. Tested: YES. Documented: YES. Developer: Ken Heafield.
30. Recaser train script updated to support IRSTLM as well. Finished coding: YES. Tested:
YES. Documented: YES. Developer: Jehan.
31. extract-ghkm. Finished coding: UNKNOWN. Tested: UNKNOWN. Documented: UNKNOWN. Developer: Philip Williams.
32. PRO tuning algorithm. Finished coding: YES. Tested: YES. Documented: YES. Developer:
50
2. Installation
33. Cruise control. Finished coding: YES. Tested: YES. Documented: YES. Developer: Ales
Tamchyna.
34. Faster SCFG rule table format. Finished coding: YES. Tested: UNKNOWN. Documented:
NO. Developer: Philip Williams.
35. LM OOV feature. Finished coding: YES. Tested: UNKNOWN. Documented: NO. Developer: Barry Haddow and Ken Heafield.
36. TER Scorer in MERT. Finished coding: UNKNOWN. Tested: UNKNOWN. Documented:
NO. Developer: Matous Machacek & Christophe Servan.
37. Multi-threading for decoder & MERT. Finished coding: YES. Tested: YES. Documented:
YES. Developer: Barry Haddow et al.
38. Expose n-gram length as part of LM state calculation.Finished coding: YES. Tested: UNKNOWN. Documented: NO. Developer: Ken Heafield and Marc Legendre.
39. Changes to chart decoder cube pruning: create one cube per dotted rule instead of one
per translation. Finished coding: YES. Tested: YES. Documented: NO. Developer: Philip
Williams.
40. Syntactic LM. Finished coding: YES. Tested: YES. Documented: YES. Developer: Lane
Schwartz.
41. Czech detokenization. Finished coding: YES. Tested: UNKNOWN. Documented: UNKNOWN. Developer: Ondrej Bojar.
2.4.6
Status 13th August, 2010
Changes since the last status report:
1. change or delete character Ø to 0 in extract-rules.cpp (Raphael and Hieu Hoang)
2.4.7
Status 9th August, 2010
Changes since the last status report:
1. Add option of retaining alignment information in the phrase-based phrase table. Decoder
loads this information if present. (Hieu Hoang & Raphael Payen)
2. When extracting rules, if the source or target syntax contains an unsupported escape
sequence (anything other than "<", ">", "&", "&apos", and "&quot") then write a warning
message and skip the sentence pair (instead of asserting).
3. In bootstrap-hypothesis-difference-significance.pl, calculates the p-value and confidence
intervals not only using BLEU, but also the NIST score. (Mark Fishel)
4. Dynamic Suffix Arrays (Abby Levenberg)
6. Continue partial translation (Ondrej Bojar and Ondrej Odchazel)
7. Bug fixes, minor bits & bobs. (Philipp Koehn, Christian Hardmeier, Hieu Hoang, Barry
Haddow, Philip Williams, Ondrej Bojar, Abbey, Mark Mishel, Lane Schwartz, Nicola
Bertoldi, Raphael, ...)
2.4.8
Status 26th April, 2010
Changes since the last status report:
2.5. Work in Progress
51
1. Synchronous CFG based decoding, a la Hiero (Chiang 2005), plus with syntax. And all
the scripts to go with it. (Thanks to Philip Williams and Hieu Hoang)
2. caching clearing in IRST LM (Nicola Bertoldi)
3. Factored Language Model. (Ondrej Bojar)
4. Fixes to lattice (Christian Hardmeier, Arianna Bisazza, Suzy Howlett)
5. zmert (Ondrej Bojar)
6. Suffix arrays (Abby Levenberg)
7. Lattice MBR and consensus decoding (Barry Haddow and Abhishek Arun)
8. Simple program that illustrates how to access a phrase table on disk from an external
program (Felipe Sánchez-Martínez)
9. Odds and sods by Raphael Payen and Sara Stymne.
2.4.9
Status 1st April, 2010
Changes since the last status report:
1. Fix for Visual Studio, and potentially other compilers (thanks to Barry, Christian, Hieu)
2. Memory leak in unique n-best fixed (thanks to Barry)
3. Makefile fix for Moses server (thanks to Barry)
2.4.10
Status 26th March, 2010
Changes since the last status report:
1. Minor bug fixes & tweaks, especially to the decoder, MERT scripts (thanks to too many
people to mention)
2. Fixes to make decoder compile with most versions of gcc, Visual studio and other compilers (thanks to Tom Hoar, Jean-Bapist Fouet).
4. Update for IRSTLM (thanks to Nicola Bertoldi and Marcello Federico)
5. Run mert on a subset of features (thanks to Nicola Bertoldi)
6. Training using different alignment models (thanks to Mark Fishel)
7. "A handy script to get many translations from Google" (thanks to Ondrej Bojar)
8. Lattice MBR (thanks to Abhishek Arun and Barry Haddow)
9. Option to compile Moses as a dynamic library (thanks to Jean-Bapist Fouet).
10. Hierarchical re-ordering model (thanks to Christian Harmeier, Sara Styme, Nadi, Marcello, Ankit Srivastava, Gabriele Antonio Musillo, Philip Williams, Barry Haddow).
11. Global Lexical re-ordering model (thanks to Philipp Koehn)
12. Experiment.perl scripts for automating the whole MT pipeline (thanks to Philipp Koehn)
2.5
Work in Progress
Refer to the website (http://www.statmt.org/moses/?n=Moses.Releases)
52
2. Installation
3
Tutorials
3.1
Phrase-based Tutorial
This tutorial describes the workings of the phrase-based decoder in Moses, using a simple
3.1.1
A Simple Translation Model
Let us begin with a look at the toy phrase-based translation model that is available for download at http://www.statmt.org/moses/download/sample-models.tgz. Unpack the tar ball and enter the directory sample-models/phrase-model.
The model consists of two files:
• phrase-table the phrase translation table, and
• moses.ini the configuration file for the decoder.
Let us look at the first line of the phrase translation table (file phrase-table):
der ||| the ||| 0.3 ||| |||
This entry means that the probality of translating the English word the from the German der
is 0.3. Or in mathematical notation: p(the|der)=0.3. Note that these translation probabilities are
in the inverse order due to the noisy channel model.
The translation tables are the main knowledge source for the machine translation decoder. The
decoder consults these tables to figure out how to translate input in one language into output
in another language.
Being a phrase translation model, the translation tables do not only contain single word entries,
but multi-word entries. These are called phrases, but this concept means nothing more than an
arbitrary sequence of words, with no sophisticated linguistic motivation.
Here is an example for a phrase translation entry in phrase-table:
53
54
3. Tutorials
das ist ||| this is ||| 0.8 ||| |||
3.1.2
Running the Decoder
Without further ado, let us run the decoder (it needs to be run from the sample-models directory) :
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini > out
Defined parameters (per moses.ini or switch):
config: phrase-model/moses.ini
input-factors: 0
lmodel-file: 8 0 3 lm/europarl.srilm.gz
mapping: T 0
n-best-list: nbest.txt 100
ttable-file: 0 0 0 1 phrase-model/phrase-table
ttable-limit: 10
weight-d: 1
weight-l: 1
weight-t: 1
weight-w: 0
Loading the LM will be faster if you build a binary file.
----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
****************************************************************************************************
The ARPA file is missing <unk>. Substituting log10 probability -100.000.
filePath: phrase-model/phrase-table
----5---10---15---20---25---30---35---40---45---50---55---60---65---70---75---80---85---90---95--100
****************************************************************************************************
IO from STDOUT/STDIN
Created input-output object : [2.000] seconds
Translating line 0 in thread id 0
Translating: das ist ein kleines haus
Collecting options took 0.000 seconds
Search took 0.000 seconds
BEST TRANSLATION: this is a small house [11111]
Translation took 0.000 seconds
Finished translating
[total=-28.923] <<0.000, -5.000, 0.000, -27.091, -1.833>>
% cat out
this is a small house
Here, the toy model managed to translate the German input sentence das ist ein kleines
haus into the English this is a small house, which is a correct translation.
3.1.
Phrase-based Tutorial
55
The decoder is controlled by the configuration file moses.ini. The file used in the example
above is displayed below.
#########################
### MOSES CONFIG FILE ###
#########################
# input factors
[input-factors]
0
# mapping steps, either (T) translation or (G) generation
[mapping]
T 0
[feature]
KENLM name=LM factor=0 order=3 num-features=1 path=lm/europarl.srilm.gz
Distortion
WordPenalty
UnknownWordPenalty
PhraseDictionaryMemory input-factor=0 output-factor=0 path=phrase-model/phrase-table num-features=1 table-limit=10
[weight]
WordPenalty0= 0
LM= 1
Distortion0= 1
PhraseDictionaryMemory0= 1
[n-best-list]
nbest.txt
100
We will take a look at all the parameters that are specified here (and then some) later. At this
point, let us just note that the translation model files and the language model file are specified
here. In this example, the file names are relative paths, but usually having full paths is better,
so that the decoder does not have to be run from a specific directory.
We just ran the decoder on a single sentence provided on the command line. Usually we want
to translate more than one sentence. In this case, the input sentences are stored in a file, one
sentence per line. This file is piped into the decoder and the output is piped into some output
file for further processing:
% moses -f phrase-model/moses.ini < phrase-model/in > out
3.1.3
Trace
How the decoder works is described in detail in the background (Section 6.1) section. But let
us first develop an intuition by looking under the hood. There are two switches that force the
decoder to reveal more about its inner workings: -report-segmentation and -verbose.
The trace option reveals which phrase translations were used in the best translation found by
the decoder. Running the decoder with the segmentation trace switch (short -t) on the same
example
56
3. Tutorials
echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -t >out
gives us the extended output
% cat out
this is |0-1| a |2-2| small |3-3| house |4-4|
Each generated English phrase is now annotated with additional information:
•
•
•
•
this is was generated from the German words 0-1, das ist,
a was generated from the German word 2-2, ein,
small was generated from the German word 3-3, kleines, and
house was generated from the German word 4-4, haus.
Note that the German sentence does not have to be translated in sequence. Here an example,
where the English output is reordered:
echo ’ein haus ist das’ | moses -f phrase-model/moses.ini -t -weight-overwrite "Distortion0= 0"
The output of this command is:
this |3-3| is |2-2| a |0-0| house |1-1|
3.1.4
Verbose
Now for the next switch, -verbose (short -v), that displays additional run time information.
The verbosity of the decoder output exists in three levels. The default is 1. Moving on to -v 2
gives additional statistics for each translated sentences:
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2
[...]
TRANSLATING(1): das ist ein kleines haus
Total translation options: 12
Total translation options pruned: 0
3.1.
Phrase-based Tutorial
57
A short summary on how many translations options were used for the translation of these
sentences.
Stack
Stack
Stack
Stack
Stack
Stack
sizes:
sizes:
sizes:
sizes:
sizes:
sizes:
1,
1,
1,
1,
1,
1,
10,
10,
10,
10,
10,
10,
2, 0, 0, 0
27, 6, 0, 0
27, 47, 6, 0
27, 47, 24, 1
27, 47, 24, 3
27, 47, 24, 3
The stack sizes after each iteration of the stack decoder. An iteration is the processing of all
hypotheses on one stack: After the first iteration (processing the initial empty hypothesis), 10
hypothesis that cover one German word are placed on stack 1, and 2 hypotheses that cover two
foreign words are placed on stack 2. Note how this relates to the 12 translation options.
total hypotheses generated = 453
number recombined = 69
number pruned = 0
During the beam search a large number of hypotheses are generated (453). Many are discarded
early because they are deemed to be too bad (272), or pruned at some later stage (0), and some
are recombined (69). The remainder survives on the stacks.
total source words = 5
words deleted = 0 ()
words inserted = 0 ()
Some additional information on word deletion and insertion, two advanced options that are
not activated by default.
BEST TRANSLATION: this is a small house [11111]
Sentence Decoding Time: : [4.000] seconds
[total=-28.923] <<0.000, -5.000, 0.000, -27.091, -1.833
And finally, the translated sentence, its coverage vector (all 5 bits for the 5 German input words
are set), its overall log-probability score, and the breakdown of the score into language model,
reordering model, word penalty and translation model components.
58
3. Tutorials
Also, the sentence decoding time is given.
The most verbose output -v 3 provides even more information. In fact, it is so much, that we
could not possibly fit it in this tutorial. Run the following command and enjoy:
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 3
Let us look together at some highlights. The overall translation score is made up from several
components. The decoder reports these components, in our case:
The score component vector looks like this:
0 distortion score
1 word penalty
2 unknown word penalty
3 3-gram LM score, factor-type=0, file=lm/europarl.srilm.gz
4 Translation score, file=phrase-table
Before decoding, the phrase translation table is consulted for possible phrase translations. For
some phrases, we find entries, for others we find nothing. Here an excerpt:
[das ; 0-0]
the , pC=-0.916, c=-5.789
this , pC=-2.303, c=-8.002
it , pC=-2.303, c=-8.076
[das ist ; 0-1]
it is , pC=-1.609, c=-10.207
this is , pC=-0.223, c=-10.291
[ist ; 1-1]
is , pC=0.000, c=-4.922
’s , pC=0.000, c=-6.116
The pair of numbers next to a phrase is the coverage, pC denotes the log of the phrase translation
probability, after c the future cost estimate for the phrase is given.
Future cost is an estimate of how hard it is to translate different parts of the sentence. After
looking up phrase translation probabilities, future costs are computed for all contigous spans
over the sentence:
3.1.
Phrase-based Tutorial
future
future
future
future
future
future
future
future
future
future
future
future
future
future
future
cost
cost
cost
cost
cost
cost
cost
cost
cost
cost
cost
cost
cost
cost
cost
from
from
from
from
from
from
from
from
from
from
from
from
from
from
from
0
0
0
0
0
1
1
1
1
2
2
2
3
3
4
to
to
to
to
to
to
to
to
to
to
to
to
to
to
to
0
1
2
3
4
1
2
3
4
2
3
4
3
4
4
is
is
is
is
is
is
is
is
is
is
is
is
is
is
is
59
-5.789
-10.207
-15.722
-25.443
-34.709
-4.922
-10.437
-20.158
-29.425
-5.515
-15.236
-24.502
-9.721
-18.987
-9.266
Some parts of the sentence are easier to translate than others. For instance the estimate for
translating the first two words (0-1: das ist) is deemed to be cheaper (-10.207) than the last
two (3-4: kleines haus, -18.987). Again, the negative numbers are log-probabilities.
After all this preperation, we start to create partial translations by translating a phrase at a time.
The first hypothesis is generated by translating the first German word as the:
creating hypothesis 1 from 0 ( <s> )
base score 0.000
covering 0-0: das
translated as: the
score -2.951 + future cost -29.425 = -32.375
unweighted feature scores: <<0.000, -1.000, 0.000, -2.034, -0.916>>
added hyp to stack, best on stack, now size 1
Here, starting with the empty initial hypothesis 0, a new hypothesis (id=1) is created. Starting
from zero cost (base score), translating the phrase das into the carries translation cost (-0.916),
distortion or reordering cost (0), language model cost (-2.034), and word penalty (-1). Recall
that the score component information is printed out earlier, so we are able to interpret the
vector.
Overall, a weighted log-probability cost of -2.951 is accumulated. Together with the future cost
estimate for the remaining part of the sentence (-29.425), this hypothesis is assigned a score of
-32.375.
And so it continues, for a total of 453 created hypothesis. At the end, the best scoring final
hypothesis is found and the hypothesis graph traversed backwards to retrieve the best translation:
60
3. Tutorials
Best path: 417 <= 285 <= 163 <= 5 <= 0
Confused enough yet? Before we get caught too much in the intricate details of the inner
workings of the decoder, let us return to actually using it. Much of what has just been said will
become much clearer after reading the background (Section 6.1) information.
3.1.5
Tuning for Quality
The key to good translation performance is having a good phrase translation table. But some
tuning can be done with the decoder. The most important is the tuning of the model parameters.
The probability cost that is assigned to a translation is a product of probability costs of four
models:
•
•
•
•
phrase translation table,
language model,
reordering model, and
word penalty.
Each of these models contributes information over one aspect of the characteristics of a good
translation:
• The phrase translation table ensures that the English phrases and the German phrases
are good translations of each other.
• The language model ensures that the output is fluent English.
• The distortion model allows for reordering of the input sentence, but at a cost: The more
reordering, the more expensive is the translation.
• The word penalty ensures that the translations do not get too long or too short.
Each of the components can be given a weight that sets its importance. Mathematically, the
cost of translation is:
p(e|f ) = φ(f |e)weightφ × LMweightLM × D(e, f )weightd × W (e)weightφ
(3.1)
The probability p(e|f) of the English translation e given the foreign input f is broken up into
four models, phrase translation phi(f|e), language model LM(e), distortion model D(e,f), and
word penalty W(e) = exp(length(e)). Each of the four models is weighted by a weight.
The weighting is provided to the decoder with the four parameters weight-t, weight-l, weight-d,
and weight-w. The default setting for these weights is 1, 1, 1, and 0. These are also the values
in the configuration file moses.ini.
Setting these weights to the right values can improve translation quality. We already sneaked
in one example above. When translating the German sentence ein haus ist das, we set the
distortion weight to 0 to get the right translation:
3.1.
Phrase-based Tutorial
61
% echo ’ein haus ist das’ | moses -f phrase-model/moses.ini -d 0
this is a house
With the default weights, the translation comes out wrong:
% echo ’ein haus ist das’ | moses -f phrase-model/moses.ini
a house is the
What is the right weight setting depends on the corpus and the language pair. Ususally, a held
out development set is used to optimize the parameter settings. The simplest method here is
to try out with a large number of possible settings, and pick what works best. Good values
for the weights for phrase translation table (weight-t, short tm), language model (weight-l,
short lm), and reordering model (weight-d, short d) are 0.1-1, good values for the word penalty
(weight-w, short w) are -3-3. Negative values for the word penalty favor longer output, positive
values favor shorter output.
3.1.6
Tuning for Speed
Let us now look at some additional parameters that help to speed up the decoder. Unfortunately higher speed usually comes at cost of translation quality. The speed-ups are achieved by
limiting the search space of the decoder. By cutting out part of the search space, we may not be
able to find the best translation anymore.
Translation Table Size
One strategy to limit the search space is by reducing the number of translation options used for
each input phrase, i.e. the number of phrase translation table entries that are retrieved. While
in the toy example, the translation tables are very small, these can have thousands of entries
per phrase in a realistic scenario. If the phrase translation table is learned from real data, it
contains a lot of noise. So, we are really interested only in the most probable ones and would
like to elimiate the others.
The are two ways to limit the translation table size: by a fixed limit on how many translation
options are retrieved for each input phrase, and by a probability threshold, that specifies that
the phrase translation probability has to be above some value.
Compare the statistics and the translation output for our toy model, when no translation table
limit is used
62
3. Tutorials
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -ttable-limit 0 -v 2
[...]
Total translation options: 12
[...]
total hypotheses generated = 453
number recombined = 69
number pruned = 0
[...]
BEST TRANSLATION: this is a small house [11111] [total=-28.923]
with the statistics and translation output, when a limit of 1 is used
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -ttable-limit 1 -v 2
[...]
Total translation options: 6
[...]
total hypotheses generated = 127
number recombined = 8
number pruned = 0
[...]
BEST TRANSLATION: it is a small house [11111] [total=-30.327]
Reducing the number of translation options to only one per phrase, had a number of effects:
(1) Overall only 6 translation options instead of 12 translation options were collected. (2) The
number of generated hypothesis fell to 127 from 442, and no hypotheses were pruned out. (3)
The translation changed, and the output now has lower log-probability: -30.327 vs. -28.923.
Hypothesis Stack Size (Beam)
A different way to reduce the search is to reduce the size of hypothesis stacks. For each number of foreign words translated, the decoder keeps a stack of the best (partial) translations.
By reducing this stack size the search will be quicker, since less hypotheses are kept at each
stage, and therefore less hypotheses are generated. This is explained in more detail on the
Background (Section 6.1) page.
From a user perspective, search speed is linear to the maximum stack size. Compare the following system runs with stack size 1000, 100 (the default), 10, and 1:
3.1.
Phrase-based Tutorial
63
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 1000
[...]
total hypotheses generated = 453
number recombined = 69
number pruned = 0
[...]
BEST TRANSLATION: this is a small house [11111] [total=-28.923]
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 100
[...]
total hypotheses generated = 453
number recombined = 69
number pruned = 0
[...]
BEST TRANSLATION: this is a small house [11111] [total=-28.923]
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 10
[...]
total hypotheses generated = 208
number recombined = 23
number pruned = 42
[...]
BEST TRANSLATION: this is a small house [11111] [total=-28.923]
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 1
[...]
total hypotheses generated = 29
number recombined = 0
number pruned = 4
[...]
BEST TRANSLATION: this is a little house [11111] [total=-30.991]
Note that the number of hypothesis entered on stacks is getting smaller with the stack size: 453,
453, 208, and 29.
As we have previously described with translation table pruning, we may also want to use the
relative scores of hypothesis for pruning instead of a fixed limit. The two strategies are also
called histogram pruning and threshold pruning.
Here some experiments to show the effects of different stack size limits and beam size limits.
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 100 -b 0
[...]
64
3. Tutorials
total hypotheses generated = 1073
number recombined = 720
number pruned = 73
[...]
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 1000 -b 0
[...]
total hypotheses generated = 1352
number recombined = 985
number pruned = 0
[...]
% echo ’das ist ein kleines haus’ | moses -f phrase-model/moses.ini -v 2 -s 1000 -b 0.1
[...]
total hypotheses generated = 45
number recombined = 3
number pruned = 0
[...]
In the second example no pruning takes place, which means an exhaustive search is performed.
With small stack sizes or small thresholds we risk search errors, meaning the generation of
translations that score worse than the best translation according to the model.
In this toy example, a worse translation is only generated with a stack size of 1. Again, by
worse translation, we mean worse scoring according to our model (-30.991 vs. -28.923). If it is
actually a worse translation in terms of translation quality, is another question. However, the
task of the decoder is to find the best scoring translation. If worse scoring translations are of
better quality, then this is a problem of the model, and should be resolved by better modeling.
3.1.7
Limit on Distortion (Reordering)
The basic reordering model implemented in the decoder is fairly weak. Reordering cost is
measured by the number of words skipped when foreign phrases are picked out of order.
Total reordering cost is computed by D(e,f) = - Σi (d_i) where d for each phrase i is defined as
d = abs( last word position of previously translated phrase + 1 - first word position of newly
translated phrase ).
This is illustrated by the following graph:
3.2.
Tutorial for Using Factored Models
65
This reordering model is suitable for local reorderings: they are discouraged, but may occur
with sufficient support from the language model. But large-scale reorderings are often arbitrary
and effect translation performance negatively.
By limiting reordering, we can not only speed up the decoder, often translation performance is
increased. Reordering can be limited to a maximum number of words skipped (maximum d)
with the switch -distortion-limit, or short -dl.
Setting this parameter to 0 means monotone translation (no reordering). If you want to allow
unlimited reordering, use the value -1.
3.2
Tutorial for Using Factored Models
Note: There may be some discrepancies between this description and the actual workings of
the training script.
• Train an unfactored model (Section 3.2.1)
• Train a model with POS tags (Section 3.2.2)
• Train a model with generation and translation steps (Section 3.2.3)
• Train a morphological analysis and generation model (Section 3.2.4)
• Train a model with multiple decoding paths (Section 3.2.5)
To work through this tutorial, you first need to have the data in place. The instructions also
assume that you have the training script and the decoder in you executable path.
You can obtain the data as follows:
• tar xzf factored-corpus.tgz
the right navigation column under "Training".
66
3.2.1
3. Tutorials
Train an unfactored model
The corpus package contains language models and parallel corpora with POS and lemma factors. Before playing with factored models, let us start with training a traditional phrase-based
model:
% train-model.perl \
--root-dir unfactored \
--corpus factored-corpus/proj-syndicate \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--external-bin-dir .../tools \
--input-factor-max 4
This creates a phrase-based model in the directory unfactored/model in about 20 minutes (on
a 2.8GHZ machine). For a quicker training run that only takes a few minutes (with much worse
results) use the just the first 1000 sentence pairs of the corpus, contained in factored-corpus/proj-syndicate.
% train-model.perl \
--root-dir unfactored \
--corpus factored-corpus/proj-syndicate.1000 \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--external-bin-dir .../tools \
--input-factor-max 4
This creates a typical phrase-based model, as specified in the created configuration file moses.ini.
Here the part of the file that points to the phrase table:
[feature]
PhraseDictionaryMemory ... path=/.../phrase-table.gz ...
You can take a look at the generated phrase table, which starts as usual with rubbish but then
occasionally contains some nice entries. The scores ensure that during decoding the good entries are preferred.
3.2.
Tutorial for Using Factored Models
67
! ||| ! ||| 1 1 1 1 2.718
" ( ||| " ( ||| 1 0.856401 1 0.779352 2.718
" ) , ein neuer film ||| " a new film ||| 1 0.0038467 1 0.128157 2.718
" ) , ein neuer film über ||| " a new film about ||| 1 0.000831718 1 0.0170876 2.71
[...]
frage ||| issue ||| 0.25 0.285714 0.25 0.166667 2.718
frage ||| question ||| 0.75 0.555556 0.75 0.416667 2.718
3.2.2
Train a model with POS tags
Take a look at the training data. Each word is not only represented by its surface form (as you
would expect in raw text), but also with additional factors.
% tail -n 1 factored-corpus/proj-syndicate.??
==> factored-corpus/proj-syndicate.de <==
korruption|korruption|nn|nn.fem.cas.sg floriert|florieren|vvfin|vvfin .|.|per|per
==> factored-corpus/proj-syndicate.en <==
corruption|corruption|nn flourishes|flourish|nns .|.|.
The German factors are
•
•
•
•
surface form,
lemma,
part of speech, and
part of speech with additional morphological information.
The English factors are
• surface form,
• lemma, and
• part of speech.
Let us start simple and build a translation model that adds only the target part-of-speech factor
on the output side:
% train-model.perl \
--root-dir pos \
--corpus factored-corpus/proj-syndicate.1000 \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--lm 2:3:factored-corpus/pos.lm:0 \
--translation-factors 0-0,2 \
--external-bin-dir .../tools
68
3. Tutorials
Here, we specify with --translation-factors 0-0,2 that the input factor for the translation
table is the (0) surface form, and the output factor is (0) surface form and (2) part of speech.
[feature]
PhraseDictionaryMemory ... input-factor=0 output-factor=0,2
The resulting phrase table looks very similar, but now also contains part-of-speech tags on the
English side:
! ||| !|. ||| 1 1 1 1 2.718
" ( ||| "|" (|( ||| 1 0.856401 1 0.779352 2.718
" ) , ein neuer film ||| "|" a|dt new|jj film|nn ||| 1 0.00403191 1 0.128157 2.718
" ) , ein neuer film über ||| "|" a|dt new|jj film|nn about|in ||| 1 0.000871765 1 0.0170876 2.718
[...]
frage ||| issue|nn ||| 0.25 0.285714 0.25 0.166667 2.718
frage ||| question|nn ||| 0.75 0.625 0.75 0.416667 2.718
We also specified two language models. Besides the regular language model based on surface
forms, we have a second language model that is trained on POS tags. In the configuration file
this is indicated by two lines in the LM section:
[feature]
SRILM name=LM0 ...
SRILM name=LM1 ...
Also, two language model weights are specified:
[weight]
LM0= 0.5
LM1= 0.5
The part-of-speech language model includes preferences such as that determiner-adjective is
likely followed by a noun, and less likely by a determiner:
-0.192859
-2.952967
dt jj nn
dt jj dt
3.2.
Tutorial for Using Factored Models
69
This model can be used just like normal phrase based models:
% echo ’putin beschreibt menschen .’ > in
% moses -f pos/model/moses.ini < in
[...]
BEST TRANSLATION: putin|nnp describes|vbz people|nns .|. [1111] [total=-6.049]
<<0.000, -4.000, 0.000, -29.403, -11.731, -0.589, -1.303, -0.379, -0.556, 4.000>>
[...]
During the decoding process, not only words (putin), but also part-of-speech are generated
(nnp).
Let’s take a look what happens, if we input a German sentence that starts with the object:
% echo ’menschen beschreibt putin .’ > in
% moses -f pos/model/moses.ini < in
BEST TRANSLATION: people|nns describes|vbz putin|nnp .|. [1111] [total=-8.030]
<<0.000, -4.000, 0.000, -31.289, -17.770, -0.589, -1.303, -0.379, -0.556, 4.000>>
Now, this is not a very good translation. The model’s aversion to do reordering trumps our
ability to come up with a good translation. If we downweight the reordering model, we get a
better translation:
% moses -f pos/model/moses.ini < in -d 0.2
BEST TRANSLATION: putin|nnp describes|vbz people|nns .|. [1111] [total=-7.649]
<<-8.000, -4.000, 0.000, -29.403, -11.731, -0.589, -1.303, -0.379, -0.556, 4.000>>
Note that this better translation is mostly driven by the part-of-speech language model, which
prefers the sequence nnp vbz nns . (-11.731) over the sequence nns vbz nnp . (-17.770).
The surface form language model only shows a slight preference (-29.403 vs. -31.289). This
is because these words have not been seen next to each other before, so the language model
has very little to work with. The part-of-speech language model is aware of the count of the
nouns involved and prefers a singular noun before a singular verb (nnp vbz) over a plural
noun before a singluar verb (nns vbz).
To drive this point home, the unfactored model is not able to find the right translation, even
with downweighted reordering model:
% moses -f unfactored/model/moses.ini < in -d 0.2
people describes putin . [1111] [total=-11.410]
<<0.000, -4.000, 0.000, -31.289, -0.589, -1.303, -0.379, -0.556, 4.000>>
70
3.2.3
3. Tutorials
Train a model with generation and translation steps
Let us now train a slightly different factored model with the same factors. Instead of mapping
from the German input surface form directly to the English output surface form and part of
speech, we now break this up into two mapping steps, one translation step that maps surface
forms to surface forms, and a second step that generates the part of speech from the surface
form on the output side:
% train-model.perl \
--root-dir pos-decomposed \
--corpus factored-corpus/proj-syndicate.1000 \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--lm 2:3:factored-corpus/pos.lm:0 \
--translation-factors 0-0 \
--generation-factors 0-2 \
--decoding-steps t0,g0 \
--external-bin-dir .../tools
Now, the translation step is specified only between surface forms (--translation-factors
0-0) and a generation step is specified (--generation-factors 0-2), mapping (0) surface form
to (2) part of speech. We also need to specified in which order the mapping steps are applied
(--decoding-steps t0,g0).
Besides the phrase table that has the same format as the unfactored phrase table, we now also
have a generation table. It is referenced in the configuration file:
[feature]
Generation ... input-factor=0 output-factor=2
[weight]
GenerationModel0= 0.3 0
Let us take a look at the generation table:
% more pos-decomposed/model/generation.0-2
nigerian nnp 1.0000000 0.0008163
proven vbn 1.0000000 0.0021142
issue nn 1.0000000 0.0021591
[...]
control vb 0.1666667 0.0014451
control nn 0.8333333 0.0017992
[...]
3.2.
Tutorial for Using Factored Models
71
The beginning is not very interesting. As most words, nigerian, proven, and issue occur only
with one part of speech, e.g., p(nnp|nigerian) = 1.0000000. Some words, however, such as
control occur with multiple part of speech, such as base form verb (vb) and single noun (nn).
The table also contains the reverse translation probability p(nigerian|nnp) = 0.0008163. In our
example, this may not be a very useful feature. It basically hurts open class words, especially
unusual ones. If we do not want this feature, we can also train the generation model as singlefeatured by the switch --generation-type single.
3.2.4
Train a morphological analysis and generation model
Translating surface forms seems to be a somewhat questionable pursuit. It does not seem
to make much sense to treat different word forms of the same lemma, such as mensch and
menschen differently. In the worst case, we will have seen only one of the word forms, so we
are not able to translate the other. This is what in fact happens in this example:
% echo ’ein mensch beschreibt putin .’ > in
% moses.1430.srilm -f unfactored/model/moses.ini < in
a mensch|UNK|UNK|UNK describes putin . [11111] [total=-158.818]
<<0.000, -5.000, -100.000, -127.565, -1.350, -1.871, -0.301, -0.652, 4.000>>
Factored translation models allow us to create models that do morphological analysis and decomposition during the translation process. Let us now train such a model:
% train-model.perl \
--root-dir morphgen \
--corpus factored-corpus/proj-syndicate.1000 \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--lm 2:3:factored-corpus/pos.lm:0 \
--translation-factors 1-1+3-2 \
--generation-factors 1-2+1,2-0 \
--decoding-steps t0,g0,t1,g1 \
--external-bin-dir .../tools
We have a total of four mapping steps:
•
•
•
•
a translation step that maps lemmas (1-1),
a generation step that sets possible part-of-speech tags for a lemma (1-2),
a translation step that maps morphological information to part-of-speech tags (3-2), and
a generation step that maps part-of-speech tag and lemma to a surface form (1,2-0).
This enables us now to translate the sentence above:
72
3. Tutorials
% echo ’ein|ein|art|art.indef.z mensch|mensch|nn|nn.masc.nom.sg \
beschreibte|beschreiben|vvfin|vvfin putin|putin|nn|nn.masc.cas.sg \
.|.|per|per’ > in
% moses -f morphgen/model/moses.ini < in
BEST TRANSLATION: a|a|dt individual|individual|nn describes|describe|vbz \
putin|putin|nnp .|.|. [11111] [total=-17.269]
<<0.000, -5.000, 0.000, -38.631, -13.357, -2.773, -21.024, 0.000, -1.386, \
-1.796, -4.341, -3.189, -4.630, 4.999, -13.478, -14.079, -4.911, -5.774, 4.999>>
Note that this is only possible, because we have seen an appropriate word form in the output
language. The word individual occurs as single noun in the parallel corpus, as translation of
einzelnen. To overcome this limitation, we may train generation models on large monolingual
corpora, where we expect to see all possible word forms.
3.2.5
Train a model with multiple decoding paths
Decomposing translation into a process of morphological analysis and generation will make
our translation model more robust. However, if we have seen a phrase of surface forms before,
it may be better to take advantage of such rich evidence.
The above model poorly translates sentences, as it does use the source surface form at all,
relying on translating the properties of the surface forms.
In practice, we fair better when we allow both ways to translate in parallel. Such a model
is trained by the introduction of decoding paths. In our example, one decoding path is the
morphological analysis and generation as above, the other path the direct mapping of surface
forms to surface forms (and part-of-speech tags, since we are using a part-of-speech tag language model):
% train-model.perl \
--corpus factored-corpus/proj-syndicate.1000 \
--root-dir morphgen-backoff \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--lm 2:3:factored-corpus/pos.lm:0 \
--translation-factors 1-1+3-2+0-0,2 \
--generation-factors 1-2+1,2-0 \
--decoding-steps t0,g0,t1,g1:t2 \
--external-bin-dir .../tools
This command is almost identical to the previous training run, except for the additional translation table 0-0,2 and its inclusion as a different decoding path :t2.
3.3.
Syntax Tutorial
73
A strategy for translating surface forms which have not been seen in the training corpus is to
translate its lemma instead. This is especially useful for translation from morphologically rich
languages to simpler languages, such as German to English translation.
% train-model.perl \
--corpus factored-corpus/proj-syndicate.1000 \
--root-dir lemma-backoff \
--f de --e en \
--lm 0:3:factored-corpus/surface.lm:0 \
--lm 2:3:factored-corpus/pos.lm:0 \
--translation-factors 0-0,2+1-0,2 \
--decoding-steps t0:t1 \
--external-bin-dir .../tools
3.3
Syntax Tutorial
24 And the people murmured against Moses, saying, What shall we drink?
25 And he cried unto the Lord; and the Lord showed him a tree, which when he had cast into the waters,
Exodus 15, 24-25
Moses supports models that have become known as hierarchical phrase-based models and syntaxbased models. These models use a grammar consisting of SCFG (Synchronous Context-Free
Grammar) rules. In the following, we refer to these models as tree-based models.
3.3.1
Tree-Based Models
Traditional phrase-based models have as atomic translation step the mapping of an input
phrase to an output phrase. Tree-based models operate on so-called grammar rules, which
include variables in the mapping rules:
ne X1 pas -> not X1 (French-English)
ate X1 -> habe X1 gegessen (English-German)
X1 of the X2 -> le X2 X1 (English-French)
The variables in these grammar rules are called non-terminals, since their occurrence indicates
that the process has not yet terminated to produce the final words (the terminals). Besides a
74
3. Tutorials
generic non-terminal X, linguistically motivated non-terminals such as NP (noun phrase) or VP
(verb phrase) may be used as well in a grammar (or translation rule set).
We call these models tree-based, because during the translation a data structure is created that
is a called a tree. To fully make this point, consider the following input and translation rules:
Input: Das Tor geht schnell auf
Rules: Das Tor -> The door
schnell -> quickly
geht X1 auf -> opens X1
X1 X2 -> X1 X2
When applying these rules in the given order, we produce the translation The door opens quickly
in the following fashion:
First the simple phrase mappings (1) Das Tor to The door and (2) schnell to quickly are
carried out. This allows for the application of the more complex rule (3) geht X1 auf to opens
X1 . Note that at this point, the non-terminal X, which covers the input span over schnell is
replaced by a known translation quickly. Finally, the glue rule (4) X1 X2 to X1 X2 combines the
two fragments into a complete sentence.
Here is how the spans over the input words are getting filled in:
|4 ---- The door opens quickly ---|
|3 --- opens quickly --|1 The door |
|2 quickly |
| Das | Tor | geht | schnell | auf
|
|
|
|
Formally, such context-free grammars are more constraint than the formalism for phrase-based
models. In practice, however, phrase-based models use a reordering limit, which leads to linear
decoding time. For tree-based models, decoding is not linear with respect to sentence length,
unless reordering limits are used.
Current research in tree-based models has the expectation to build translation models that more
closely model the underlying linguistic structure of language, and its essential element: recursion. This is an active field of research.
3.3.
Syntax Tutorial
75
A Word on Terminology
You may have read in the literature about hierarchical phrase-based, string-to-tree, tree-tostring, tree-to-tree, target-syntactified, syntax-augmented, syntax-directed, syntax-based, grammarbased, etc., models in statistical machine translation. What do the tree-based models support?
All of the above.
The avalanche of terminology stems partly from the need of researchers to carve out their own
niche, partly from the fact that work in this area has not yet fully settled on a agreed framework,
but also from a fundamental difference. As we already pointed out, the motivation for treebased models are linguistic theories and their syntax trees. So, when we build a data structure
called a tree (as Computer Scientist call it), do we mean that we build a linguistic syntax tree (as
Linguists call it)?
Not always, and hence the confusion. In all our examples above we used a single non-terminal
X, so not many will claim the the result is a proper linguistic syntax with its noun phrases NP,
verb phrases VP, and so on. To distinguish models that use proper linguistic syntax on the input
side, on the output side, on both, or on neither all this terminology has been invented.
Let’s decipher common terms found in the literature:
•
•
•
•
•
•
•
•
•
hierarchical phrase-based: no linguistic syntax,
string-to-tree: linguistic syntax only in output language,
tree-to-string: linguistic syntax only in input language,
tree-to-tree: linguistic syntax in both languages,
target-syntactified: linguistic syntax only in output language,
syntax-augmented: linguistic syntax only in output language,
syntax-directed: linguistic syntax only in input language,
syntax-based: unclear, we use it for models that have any linguistic syntax, and
grammar-based: wait, what?
In this tutorial, we refer to un-annotated trees as trees, and to trees with syntactic annotation
as syntax. So a so-called string-to-tree model is here called a target-syntax model.
Chart Decoding
Phrase-Based decoding generates a sentence from left to right, by adding phrases to the end of
a partial translation. Tree-based decoding builds a chart, which consists of partial translation
for all possible spans over the input sentence.
Currently Moses implements a CKY+ algorithm for arbitrary number of non-terminals per rule
and an arbitrary number of types of non-terminals in the grammar.
3.3.2
Decoding
We assume that you have already installed the chart decoder, as described in the Get Started1
section.
You can find an example model for the decoder from the Moses web site2 . Unpack the tar ball
and enter the directory sample-models:
1
2
http://www.statmt.org/moses/?n=Development.GetStarted#chart
76
3. Tutorials
% tar xzf sample-models.tgz
% cd sample-models/string-to-tree
The decoder is called just as for phrase models:
% echo ’das ist ein haus’ | moses_chart -f moses.ini > out
% cat out
this is a house
What happened here?
Trace
Using the option -T we can some insight how the translation was assembled:
41 X TOP -> <s> S </s> (1,1)
[0..5] -3.593 <<0.000, -2.606, -9.711, 2.526>> 20
20 X S
-> NP V NP
(0,0) (1,1) (2,2) [1..4] -1.988 <<0.000, -1.737, -6.501, 2.526>> 3 5 11
3 X NP -> this
[1..1] 0.486 <<0.000, -0.434, -1.330, 2.303>>
5 X V
-> is
[2..2] -1.267 <<0.000, -0.434, -2.533, 0.000>>
11 X NP -> DT NN
(0,0) (1,1)
[3..4] -2.698 <<0.000, -0.869, -5.396, 0.000>> 7 9
7 X DT -> a
[3..3] -1.012 <<0.000, -0.434, -2.024, 0.000>>
9 X NN -> house
[4..4] -2.887 <<0.000, -0.434, -5.774, 0.000>>
Each line represents a hypothesis that is part of the derivation of the best translation. The pieces
of information in each line (with the first line as example) are:
•
•
•
•
•
•
•
•
the hypothesis number, a sequential identifier (41),
the input non-terminal (X),
the output non-termial (S),
the rule used to generate this hypothesis (TOP -> <s> S </s>),
alignment information between input and output non-terminals in the rule ((1,1)),
the span covered by the hypothesis, as defined by input word positions ([0..5]),
the score of the hypothesis (3.593),
its component scores (<<...>>):
– unknown word penalty (0.000),
– word penalty (-2.606),
– language model score (-9.711),
– rule application probability (2.526), and
• prior hypotheses, i.e. the children nodes in the tree, that this hypothesis is built on (20).
As you can see, the model used here is a target-syntax model, It uses linguistic syntactic annotation on the target side, but on the input side everything is labeled X.
3.3.
Syntax Tutorial
77
Rule Table
If we look at the string-to-tree directory, we find two files: the configuration file moses.ini
which points to the language model (in lm/europarl.srilm.gz), and the rule table file rule-table.
The configuration file moses.ini has a fairly familiar format. It is mostly identical to the configuration file for phrase-based models. We will describe further below in detail the new parameters of the chart decoder.
The rule table rule-table is an extension of the Pharaoh/Moses phrase-table, so it will be
familiar to anybody who has used it before. Here are some lines as example:
gibt [X] ||| gives [ADJ] ||| 1.0 ||| ||| 3 5
es gibt [X] ||| there is [ADJ] ||| 1.0 ||| ||| 2 3
[X][DT] [X][NN] [X] ||| [X][DT] [X][NN] [NP] ||| 1.0 ||| 0-0 1-1 ||| 2 4
[X][DT] [X][ADJ] [X][NN] [X] ||| [X][DT] [X][ADJ] [X][NN] [NP] ||| 1.0 ||| 0-0 1-1 2-2 ||| 5 6
[X][V] [X][NP] [X] ||| [X][V] [X][NP] [VP] ||| 1.0 ||| 0-0 1-1 ||| 4 3
Each line in the rule table describes one translation rule. It consists of five components separated by three bars:
•
•
•
•
•
the source string and source left-hand-side,
the target string and target left-hand-side,
the alignment between non-terminals (using word positions),
score(s): here only one, but typically multiple scores are used, and
frequency counts of source & target phrase (for debuggin purposes, not used during decoding).
The format is slightly different from the Hiero format. For example, the Hiero rule
[X] ||| [X,1] trace ’ ||| [X,1] &#52628;&#51201; ’ \
||| 0.727273 0.444625 1 0.172348 2.718
is formatted as
[X][X] trace ’ [X] ||| [X][X] &#52628;&#51201; ’ [X] \
||| 0.727273 0.444625 1 0.172348 2.718 ||| 0-0 ||| 2 3
A syntax rule in a string-to-tree grammar:
[NP] ||| all [NN,1] ||| &#47784;&#46304; [NN,1] \
||| 0.869565 0.627907 0.645161 0.243243 2.718
78
3. Tutorials
is formatted as
all [X][NN] [X] ||| &#47784;&#46304; [X][NN] [NP] \
||| 0.869565 0.627907 0.645161 0.243243 2.718 ||| 1-1 ||| 23 31
The format can also a represent a tree-to-string rule, which has no Hiero equivalent:
all [NN][X] [NP] ||| &#47784;&#46304; [NN][X] [X] \
||| 0.869565 0.627907 0.645161 0.243243 2.718 ||| 1-1 ||| 23 31
Usually, you will also need these ’glue’ rules:
<s> [X][S] </s> [X] ||| <s> [X][S] </s> [TOP] ||| 1.0 ||| 1-1
<s> [X][NP] </s> [X] ||| <s> [X][NP] </s> [TOP] ||| 1.0 ||| 1-1
<s> [X] ||| <s> [S] ||| 1 |||
[X][S] </s> [X] ||| [X][S] </s> [S] ||| 1 ||| 0-0
[X][S] [X][X] [X] ||| [X][S] [X][X] [S] ||| 2.718 ||| 0-0 1-1
Finally, this rather technical rule applies only to spans that cover everything except the sentence
boundary markers <s> and </s>. It completes a translation with of a sentence span (S).
More Example
The second rule in the table, that we just glanced at, allows something quite interesting: the
translation of a non-contiguous phrase: macht X auf.
Let us try this with the decoder on an example sentence:
% echo ’er macht das tor auf’ | moses_chart -f moses.ini -T trace-file ; cat trace-file
[...]
14 X TOP -> <s> S </s> (1,1)
[0..6] -7.833 <<0.000, -2.606, -17.163, 1.496>> 13
13 X S
-> NP VP
(0,0) (1,1) [1..5] -6.367 <<0.000, -1.737, -14.229, 1.496>> 2 11
2 X NP -> he
[1..1] -1.064 <<0.000, -0.434, -2.484, 0.357>>
11 X VP -> opens NP
(1,1)
[2..5] -5.627 <<0.000, -1.303, -12.394, 1.139>> 10
10 X NP -> DT NN
(0,0) (1,1) [3..4] -3.154 <<0.000, -0.869, -7.224, 0.916>> 6 7
6 X DT -> the
[3..3] 0.016 <<0.000, -0.434, -0.884, 0.916>>
7 X NN -> gate
[4..4] -3.588 <<0.000, -0.434, -7.176, 0.000>>
he opens the gate
3.3.
Syntax Tutorial
79
You see the creation application of the rule in the creation of hypothesis 11. It generates opens
NP to cover the input span [2..5] by using hypothesis 10, which coveres the span [3..4].
Note that this rule allows us to do something that is not possible with a simple phrase-based
model. Phrase-based models in Moses require that all phrases are contiguous, they can not
have gaps.
The final example illustrates how reordering works in a tree-based model:
% echo ’ein haus ist das’ | moses_chart -f moses.ini -T trace-file ; cat trace-file
41 X TOP -> <s> S </s> (1,1)
[0..5] -2.900 <<0.000, -2.606, -9.711, 3.912>> 18
18 X S
-> NP V NP
(0,2) (1,1) (2,0) [1..4] -1.295 <<0.000, -1.737, -6.501, 3.912>> 11 5 8
11 X NP -> DT NN
(0,0) (1,1)
[1..2] -2.698 <<0.000, -0.869, -5.396, 0.000>> 2 4
2 X DT -> a
[1..1] -1.012 <<0.000, -0.434, -2.024, 0.000>>
4 X NN -> house
[2..2] -2.887 <<0.000, -0.434, -5.774, 0.000>>
5 X V
-> is
[3..3] -1.267 <<0.000, -0.434, -2.533, 0.000>>
8 X NP -> this
[4..4] 0.486 <<0.000, -0.434, -1.330, 2.303>>
this is a house
The reordering in the sentence happens when hypothesis 18 is generated. The non-lexical rule
S -> NP V NP takes the underlying children nodes in inverse order ((0,2) (1,1) (2,0)).
Not any arbitrary reordering is allowed -– as this can be the case in phrase models. Reordering
has to be motivated by a translation rule. If the model uses real syntax, there has to be a
syntactic justification for the reordering.
3.3.3
Decoder Parameters
The most important consideration in decoding is a speed/quality trade-off. If you want to
win competitions, you want the best quality possible, even if it takes a week to translate 2000
sentences. If you want to provide an online service, you know that users get impatient, when
they have to wait more than a second.
Beam Settings
The chart decoder has an implementation of CKY decoding using cube pruning. The latter
means that only a fixed number of hypotheses are generated for each span. This number can be
changed with the option cube-pruning-pop-limit (or short cbp). The default is 1000, higher
numbers slow down the decoder, but may result in better quality.
Another setting that directly affects speed is the number of rules that are considered for each
input left hand side. It can be set with ttable-limit.
80
3. Tutorials
Limiting Reordering
The number of spans that are filled during chart decoding is quadratic with respect to sentence
length. But it gets worse. The number of spans that are combined into a span grows linear with
sentence length for binary rules, quadratic for trinary rules, and so on. In short, long sentences
become a problem. A drastic solution is the size of internal spans to a maximum number.
This sounds a bit extreme, but does make some sense for non-syntactic models. Reordering is
limited in phrase-based models, and non-syntactic tree-based models (better known as hierarchical phrase-based models) and should limit reordering for the same reason: they are just not
very good at long-distance reordering anyway.
The limit on span sizes can be set with max-chart-span. In fact its default is 10, which is not a
useful setting for syntax models.
Handling Unknown Words
In a target-syntax model, unknown words that just copied verbatim into the output need to
get a non-terminal label. In practice unknown words tend to be open class words, most likely
names, nouns, or numbers. With the option unknown-lhs you can specify a file that contains
pairs of non-terminal labels and their probability per line.
Optionally, we can also model the choice of non-terminal for unknown words through sparse
features, and optimize their cost through MIRA or PRO. This is implemented by relaxing the label matching constraint during decoding to allow soft matches, and allowing unknown words
to expand to any non-terminal. To activate this feature:
use-unknown-word-soft-matches = true (in EMS config)
-unknown-word-label FILE1 -unknown-word-soft-matches FILE2 (in train-model.perl)
Technical Settings
The parameter non-terminals is used to specify privileged non-terminals. These are used for
unknown words (unless there is a unknown word label file) and to define the non-terminal
label on the input side, when this is not specified.
Typically, we want to consider all possible rules that apply. However, with a large maximum
phrase length, too many rule tables and no rule table limit, this may explode. The number of
rules considered can be limited with rule-limit. Default is 5000.
3.3.4
Training
In short, training uses the identical training script as phrase-based models. When running
train-model.perl, you will have to specify additional parameters, e.g. -hierarchical and
-glue-grammar. You typically will also reduce the number of lexical items in the grammar with
-max-phrase-length 5.
That’s it.
3.3.
Syntax Tutorial
81
Training Parameters
There are a number of additional decisions about the type of rules you may want to include in
your model. This is typically a size / quality trade-off: Allowing more rule types increases the
size of the rule table, but lead to better results. Bigger rule tables have a negative impact on
memory use and speed of the decoder.
There are two parts to create a rule table: the extraction of rules and the scoring of rules. The
first can be modified with the parameter --extract-options="..." of train-model.perl.
The second with --score-options="...".
Here are the extract options:
• --OnlyDirect: Only creates a model with direct conditional probabilities p(f|e) instead of
the default direct and indirect (p(f|e) and p(e|f)).
• --MaxSpan SIZE: maximum span size of the rule. Default is 15.
• --MaxSymbolsSource SIZE and --MaxSymbolsTarget SIZE: While a rule may be extracted
from a large span, much of it may be knocked out by sub-phrases that are substituted by
non-terminals. So, fewer actual symbols (non-terminals and words remain). The default
maximum number of symbols is 5 for the source side, and practically unlimited (999) for
the target side.
• --MinWords SIZE: minimum number of words in a rule. Default is 1, meaning that each
rule has to have at least one word in it. If you want to allow non-lexical rules set this to
zero. You will not want to do this for hierarchical models.
• --AllowOnlyUnalignedWords: This is related to the above. A rule may have words in it,
but these may be unaligned words that are not connected. By default, at least one aligned
word is required. Using this option, this requirement is dropped.
• --MaxNonTerm SIZE: the number of non-terminals on the right hand side of the rule. This
has an effect on the arity of rules, in terms of non-terminals. Default is to generate only
binary rules, so the setting is 2.
• --MinHoleSource SIZE and --MinHoleTarget SIZE: When sub-phrases are replaced by
non-terminals, we may require a minimum size for these sub-phrases. The default is 2 on
the source side and 1 (no limit) on the target side.
• --DisallowNonTermConsecTarget and --NonTermConsecSource. We may want to restrict if there can be neighboring non-terminals in rules. In hierarchical models there
is a bad effect on decoding to allow neighboring non-terminals on the source side. The
default is to disallow this -- it is allowed on the target side. These switches override the
defaults.
• --NoFractionalCounting: For any given source span, any number of rules can be generated. By default, fractional counts are assigned, so probability of these rules adds up to
one. This option leads to the count of one for each rule.
• --NoNonTermFirstWord: Disallows that a rule starts with a non-terminal.
Once rules are collected, the file of rules and their counts have to be converted into a probabilistic model. This is called rule scoring, and there are also some additional options:
• --OnlyDirect: only estimates direct conditional probabilities. Note that this option needs
to be specified for both rule extraction and rule scoring.
• --NoLex: only includes rule-level conditional probabilities, not lexical scores.
• --GoodTuring: Uses Good Turing discounting to reduce actual accounts. This is a good
thing, use it.
82
3. Tutorials
Training Syntax Models
Training hierarchical phrase models, i.e., tree-based models without syntactic annotation, is
pretty straight-forward. Adding syntactic labels to rules, either on the source side or the target
side, is not much more complex. The main hurdle is to get the annotation. This requires a
syntactic parser.
Syntactic annotation is provided by annotating all the training data (input or output side, or
both) with syntactic labels. The format that is used for this uses XML markup. Here an example:
<tree label="NP"> <tree label="DET"> the </tree> \
<tree label="NN"> cat </tree> </tree>
So, constituents are surrounded by an opening and a closing <tree> tag, and the label is
provided with the parameter label. The XML markup also allows for the placements of the
tags in other positions, as long as a span parameter is provided:
<tree label="NP" span="0-1"/> <tree label="DET" span="0-0"/> \
<tree label="NN" span="1-1"/> the cat
After annotating the training data with syntactic information, you can simply run train-model.perl
as before, except that the switches --source-syntax or --target-syntax (or both) have to be
set.
You may also change some of the extraction settings, for instance --MaxSpan 999.
Annotation Wrappers
To obtain the syntactic annotation, you will likely use a third-party parser, which has its own
idiosyncratic input and output format. You will need to write a wrapper script that converts it
into the Moses format for syntax trees.
We provide wrappers (in scripts/training/wrapper) for the following parsers.
• Bitpar is available from the web site of the University of Munich3 . The wrapper is
parse-de-bitpar.perl
• Collins parser is availble from MIT4 . The wrapper is parse-en-collins.perl
If you wrote your own wrapper for a publicly available parsers, please share it with us!
3
4
http://www.cis.uni-muenchen.de/~schmid/tools/BitPar/
http://people.csail.mit.edu/mcollins/code.html
3.3.
Syntax Tutorial
83
Relaxing Parses
The use of syntactic annotation puts severe constraints on the number of rules that can be
extracted, since each non-terminal has to correspond to an actual non-terminal in the syntax
tree.
Recent research has proposed a number of relaxations of this constraint. The program relax-parse
(in training/phrase-extract) implements two kinds of parse relaxations: binarization and a
method proposed under the label of syntax-augmented machine translation (SAMT) by Zollmann and Venugopal.
Readers familiar with the concept of binarizing grammars in parsing, be warned: We are talking here about modifying parse trees, which changes the power of the extracted grammar, not
binarization as a optimization step during decoding.
The idea is the following: If the training data contains a subtree such as
then it is not possible to extract translation rules for Ariel Sharon without additional syntactic
context. Recall that each rule has to match a syntactic constituent.
The idea of relaxing the parse trees is to add additional internal nodes that makes the extraction of additional rules possible. For instance left-binarization adds two additional nodes and
converts the subtree into:
The additional node with the label N̂P allows for the straight-forward extraction of a translation
rule (of course, unless the word alignment does not provide a consistent alignment).
The program relax-parse allows the following tree transformations:
• --LeftBinarize and --RightBinarize: Adds internal nodes as in the example above.
Right-binarization creates a right-branching tree.
• --SAMT 1: Combines pairs of neighboring children nodes into tags, such as DET+ADJ. Also
nodes for everything except the first child (NP
DET) and everything except the last child (NP/NN) are added.
84
3. Tutorials
• --SAMT 2: Combines any pairs of neighboring nodes, not only children nodes, e.g., VP+DET.
• --SAMT 3: not implemented.
• --SAMT 4: As above, but in addition each previously unlabeled node is labeled as FAIL,
so no syntactic constraint on grammar constraint remains.
Note that you can also use both --LeftBinarize and --RightBinarize. Note that in this case,
as with all the SAMT relaxations, the resulting annotation is not any more a tree, since there is
not a single set of rule applications that generates the structure (now called a forest).
Here an example, what parse relaxation does to the number of rules extracted (English-German
News Commentary, using Bitpar for German, no English syntax):
Relaxation Setting
no syntax
basic syntax
left-binarized
right-binarized
SAMT 1
SAMT 2
SAMT 4
Number of Rules
59,079,493
2,291,400
2,914,348
2,979,830
8,669,942
35,164,756
131,889,855
On-Disk Rule Table
The rule table may become too big to fit into the RAM of the machine. Instead of loading
the rules into memory, it is also possible to leave the rule table on disk, and retrieve rules on
demand.
This is described in On-Disk Phrase Table5 .
3.3.5
Using Meta-symbols in Non-terminal Symbols (e.g., CCG)
Often a syntactic formalism will use symbols that are part of the meta-symbols that denote nonterminal boundaries in the SCFG rule table, and glue grammar. For example, in Combinatory
Categorial Grammar (CCG, Steedman, 2000), it is customary to denote grammatical features
by placing them after the non-terminal symbol inside square brackets, as in S[dcl] (declarative
sentence) vs. S[q] (interrogative sentence).
Although such annotations may be useful to discriminate good translations from bad, including square brackets in the non-terminal symbols themselves can confuse Moses. Some users
have reported that category symbols were mangled (by splitting them at the square brackets)
after converting to an on-disk representation (and potentially in other scenarios -- this is currently an open issue). A way to side-step this issue is to escape square brackets with a symbol
that is not part of the meta-language of the grammar files, e.g. using the underscore symbol:
S[dcl] => S_dcl_
and
S[q] => S_q_
5
3.3.
Syntax Tutorial
85
before extracting a grammar. This should be done in all data or tables that mention such syntactic categories. If the rule table is automatically extracted, it suffices to escape the categories
in the <tree label="..."...> mark-up that is supplied to the training script. If you roll your own
rule tables (or use an unknown-lhs file), you should make sure they are properly escaped.
3.3.6
Different Kinds of Syntax Models
String-to-Tree
Most SCFG-based machine translation decoders at the current time are designed to uses hierarchical phrase-based grammar (Chiang, 2005) or syntactic grammar. Joshua, cdec, Jane are some
of the open-sourced systems that have such decoders.
The hierarchical phrase-based grammar is well described elsewhere so we will not go into
details here. Briefly, the non-terminals are not labelled with any linguistically-motivated labels.
By convention, non-terminals have been simply labelled as X, e.g.
X --> der X1 ||| the X1
Usually, a set of glue rules are needed to ensure that the decoder always output an answer. By
convention, the non-terminals for glue rules are labelled as S, e.g.
S --> <s> ||| <s>
S --> X1 </s> ||| X1 </s>
S --> X1 X2 ||| X1 X2
In a syntactic model, non-terminals are labelled with linguistically-motivated labels such as
’NOUN’, ’VERB’ etc. For example:
DET --> der ||| the
These labels are typically obtained by parsing the target side of the training corpus. (However,
it is also possible to use parses of the source side which has been projected onto the target side
(Ambati and Chen, 2007) ).
The input to the decoder when using this model is a conventional string, as in phrase-based and
hierarchical phrase-based models. The output is a string. However, the CFG-tree derivation of
the output (target) can also be obtained (in Moses by using the -T argument), the non-terminals
in this tree will be labelled with the linguistically-motivated labels.
86
3. Tutorials
For these reasons, these syntactic models are called ’target’ syntax models, or ’string-to-tree’
model, by many in the Moses community and elsewhere. (Some papers by people at ISI inverted this naming convention due to their adherance to the noisy-channel framework).
The implementation of string-to-tree models is fairly standard and similar across different
open-source decoders such as Moses, Joshua, cdec and Jane.
The input to the model is the string:
das ist ein kleines haus
The output string is
this is a small house
The target tree it produces is
(TOP <s> (S (NP this) (VP (V is) (NP (DT a) (ADJ small) (NN house)))) </s>)
RECAP - The input is a string, the output is a tree with linguistically-motivated labels.
Tree-to-string
Unlike the string-to-tree model, the tree-to-string model is not as standardized across different
decoders. This section describes the Moses implementation.
Input tree representation The input to the decoder is a parse tree, not a string. For Moses,
the parse tree should be formatted using XML. The decoder converts the parse tree into an
annotated string (a chart?). Each span in the chart is labelled with the non-terminal from the
parse tree. For example, the input
<tree label="NP"> <tree label="DET"> the </tree> <tree label="NN"> cat </tree> </tree>
is converted to an annotated string
6
3.3.
Syntax Tutorial
87
the
cat
-DET- -NN-----NP-----
To support easier glue rules, the non-terminal ’X’ is also added for every span in the annotated
string. Therefore, the input above is actually converted to:
the
cat
-DET- -NN---X-- --X-----NP---------X-----
Translation rules During decoding, the non-terminal of the rule that spans a substring in the
sentence must match the label on the annoated string. For example, the following rules can be
applied to the above sentence.
NP
NP
NP
NP
-->
-->
-->
-->
the katze ||| die katze
the NN1 ||| der NN1
DET1 cat ||| DET1 katze
DET1 NN2 ||| DET1 NN2
However, these rules can’t as they don’t match one or more non-terminals.
VB --> the katze ||| die katze
Therefore, non-terminal in the translation rules in a tree-to-string model acts as constraints on
which rules can be applied. This constraint is in addition to the usual role of non-terminals.
A feature which is currently unique to the Moses decoder is the ability to separate out these
two roles. Each non-terminal in all translation rules is represented by two labels:
1. The source non-terminal which constrains rules to the input parse tree
2. The target non-terminal which has the normal parsing role.
88
3. Tutorials
When we need to differentiate source and target non-terminals, the translation rules are instead
written like this:
NP --> the NN1 ||| X --> der X1
This rule indicates that the non-terminal should span a NN constituent in the input text, and
that the whole rule should span an NP constituent. The target non-terminals in this rule are
both X, therefore, this rule would be considered part of tree-to-string grammar.
(Using this notation is probably wrong as the source sentence is not properly parsed - see next
section. It may be better to express the Moses tree-to-string grammar as a hierarchical grammar,
X --> the X1 ||| der X1 ||| LHS = NP, X_1 = NN
However, this may be even more confusing so we will stick with our convention for now.)
RECAP - Grammar rules in Moses have 2 labels for each non-terminals; one to constrain the
non-terminal to the input parse tree, the other is used in parsing.
Consequences
1. The Moses decoder always checks the source non-terminal, even when it is decoding with a
string-to-string or string-to-tree grammar. For example, when checking whether the following
rule can be applied
X --> der X1 ||| the X1
the decoder will check whether the RHS non-terminal, and the whole rule, spans an input parse
constituent X. Therefore, even when decoding with a string-to-string or string-to-tree grammar,
it is necessary to add the X non-terminal to every input span. For example, the input string the
cat must be annotated as follows
the
cat
--X-- --X------X-----
3.3.
Syntax Tutorial
89
to allow the string to be decoded with a string-to-string or string-to-tree grammar.
2. There is no difference between a linguistically derived non-terminal label, such as NP, VP
etc, and the non-linguistically motivated X label. They can both be used in one grammar, or
even 1 translation rule. This ’mixed-syntax’ model was explored in (Hoang and Koehn, 2010)
and in Hieu Hoang’s thesis7
3. The source non-terminals in translation rules are used just to constrain against the input
parse tree, not for parsing. For example, if the input parse tree is
(VP (NP (PRO he)) (VB goes))
and tree-to-string rules are:
PRO --> he ||| X --> il
VB --> goes ||| X --> va
VP --> NP1 VB2 ||| X --> X1 X2
This will create a valid translation. However, the span over the word ’he’ will be labelled as
PRO by the first rule, and NP by the 3rd rule. This is illustrated in more detail in Hieu’s thesis
Section 4.2.11.
4. To avoid the above and ensure that source spans are always consistently labelled, simply
project the non-terminal label to both source and target. For example, change the rule
VP --> NP1 VB2 ||| X --> X1 X2
to
VP --> NP1 VB2 ||| VP --> NP1 VB2
3.3.7
Format of text rule table
The format of the Moses rule table is different from that used by Hiero, Joshua and cdec, and
has often been a source of confusion. We shall attempt to explain the reasons in this section.
The format is derived from the Pharaoh/Moses phrase-based format. In this format, a translation rule
7
http://www.statmt.org/~s0565741/ddd.pdf
90
3. Tutorials
a b c --> d e f , with word alignments a1, a2 ..., and probabilities p1, p2, ...
is formatted as
a b c ||| d e f ||| p1 p2 ... ||| a1 a2 ...
For a hierarchical pb rule,
X --> a X1 b c X2 ||| d e f X2 X1
The Hiero/Joshua/cdec format is
X ||| a [X,1] b c [X,2] ||| d e f [X,2] [X,1] ||| p1 p2 ...
The Moses format is
a [X][X] b c [X][X] [X] ||| d e f [X][X] [X][X] [X] ||| p1 p2 ... ||| 1-3 4-4
For a string-to-tree rule,
VP --> a X1 b c X2 ||| d e f NP2 ADJ1
the Moses format is
a [X][ADJ] b c [X][NP] [X] ||| d e f [X][NP] [X][ADJ] [VP] ||| p1 p2 ... ||| 1-3 4-4
For a tree-to-string rule,
3.4.
Optimizing Moses
91
VP --> a ADJ1 b c NP2 ||| X --> d e f X2 X1
The Moses format is
a [ADJ][X] b c [NP][X] [VP] ||| d e f [NP][X] [ADJ][X] [X] ||| p1 p2 ... ||| 1-3 4-4
The main reasons for the difference between the Hiero/Joshua/cdec and Moses formats are as
follows:
1. The text rule table should be easy to convert to a binary, on-disk format. We have seen in
the community that this allows much large models to be used during decoding, even on
memory limited servers. To make the conversion efficient, the text rule table must have
the following properties:
(a) For every rule, the sequence of terminals and non-terminals in the first column (the
’source’ column) should match the lookup sequence that the decoder will perform.
(b) The file can be sorted so that the first column is in alphabetical order. The decoder
needs to look up the target non-terminals on the RHS of each rule so the first column
consist of source terminals and non-terminal, AND target non-terminals from the
RHS.
2. The phrase probability calculations should be performed efficiently. To calculate p(t | s) =
count(t,s) / count(s) the extract file must be sorted in contiguous order so that each count
can be performed and used to calculate the probability, then discarded immediately to
save memory. Similarly for p(s | t) = count(t,s) / count(t)
The Hiero/Joshua/cdec file format is sufficient for hierarchical models, but not for the various
syntax models supported by Moses.
3.4
Optimizing Moses
3.4.1
How much memory do I need during decoding?
The single-most important thing you need to run Moses fast is MEMORY. Lots of MEMORY.
(For example, the Edinburgh group have servers with 144GB of RAM). The rest of this section
is just details of how to make the training and decoding run fast.
Calculate total file size of the binary phrase tables, binary language models and binary reordering models.
For example,
92
3. Tutorials
% ll -h phrase-table.0-0.1.1.binphr.*
-rw-r--r-- 1 s0565741 users 157K 2012-06-13 12:41 phrase-table.0-0.1.1.binphr.idx
-rw-r--r-- 1 s0565741 users 5.4M 2012-06-13 12:41 phrase-table.0-0.1.1.binphr.srctree
-rw-r--r-- 1 s0565741 users 282K 2012-06-13 12:41 phrase-table.0-0.1.1.binphr.srcvoc
-rw-r--r-- 1 s0565741 users 1.1G 2012-06-13 12:41 phrase-table.0-0.1.1.binphr.tgtdata
-rw-r--r-- 1 s0565741 users 1.7M 2012-06-13 12:41 phrase-table.0-0.1.1.binphr.tgtvoc
% ll -h reordering-table.1.wbe-msd-bidirectional-fe.binlexr.*
-rw-r--r-- 1 s0565741 users 157K 2012-06-13 13:36 reordering-table.1.wbe-msd-bidirectional-fe.binlexr.idx
-rw-r--r-- 1 s0565741 users 1.1G 2012-06-13 13:36 reordering-table.1.wbe-msd-bidirectional-fe.binlexr.srctree
-rw-r--r-- 1 s0565741 users 1.1G 2012-06-13 13:36 reordering-table.1.wbe-msd-bidirectional-fe.binlexr.tgtdata
-rw-r--r-- 1 s0565741 users 282K 2012-06-13 13:36 reordering-table.1.wbe-msd-bidirectional-fe.binlexr.voc0
-rw-r--r-- 1 s0565741 users 1.7M 2012-06-13 13:36 reordering-table.1.wbe-msd-bidirectional-fe.binlexr.voc1
% ll -h interpolated-binlm.1
-rw-r--r-- 1 s0565741 users 28G 2012-06-15 11:07 interpolated-binlm.1
The total size of these files is approx. 31GB. Therefore, a translation system using these models
requires 31GB (+ roughly 500MB) of memory to run fast.
I’ve got this much memory but it’s still slow. Why?
Run this:
cat phrase-table.0-0.1.1.binphr.* > /dev/null
cat reordering-table.1.wbe-msd-bidirectional-fe.binlexr.* > /dev/null
cat interpolated-binlm.1 > /dev/null
This forces the operating system to cache the binary models in memory, minimizing pages
faults while the decoder is running. Other memory-intensive processes on the computer should
not be running, otherwise the file-system cache may be reduced.
Use huge pages
Moses does a lot of random lookups. If you’re running Linux, check that transparent huge
pages8 are enabled. If
cat /sys/kernel/mm/transparent_hugepage/enabled
responds with
8
https://lwn.net/Articles/423584/
3.4.
Optimizing Moses
93
then transparent huge pages are enabled.
On some RedHat/Centos systems, the file is /sys/kernel/mm/redhat_transparent_hugepage/enabled
and madvise will not appear. If neither file exists, upgrade the kernel to at least 2.6.38 and compile with CONFIG_SPARSEMEM_VMEMMAP. If the file exists, but the square brackets are not around
"always", then run
echo always > /sys/kernel/mm/transparent_hugepage/enabled
as root (NB: to use sudo, quote the > character). This setting will not be preserved across
reboots, so consider adding it to an init script.
Use the compact phrase and reordering table representations to reduce memory usage by a
factor of 10
See the manual on binarized9 and compact10 phrase table for a description how to compact
your phrase tables. All the things said above for the standard binary phrase table are also true
for the compact versions. The principle is the same, the total size of the binary files determines
your memory usage, but since the combined size of the compact phrase table and the compact
reordering model maybe up to 10 to 12 times smaller than with the original binary implementations, you will save exactly this much memory. You can also use the --minphr-memory and
--minlexr-memory options to load the tables into memory at Moses start-up instead of doing
the above mentioned caching trick. This may take some time during warm-up, but may save
a lot of time in the long term. If you are concerned for performance, see Junczys-Dowmunt
(2012)11 for a comparison. There is virtually no overhead due to on-the-fly decompression on
large-memory-systems and considerable speed-up on systems with limited memory.
3.4.2
How little memory can I get away with during decoding?
The decoder can run on very little memory, about 200-300MB for phrase-based and 400-500MB
for hierarchical decoding (according to Hieu). The decoder can run on an iPhone! And laptops.
However, it will be VERY slow, unless you have very small models or the models are on fast
disks such as flash disks.
9
11
http://ufal.mff.cuni.cz/pbml/98/art-junczys-dowmunt.pdf
10
94
3. Tutorials
3.4.3
Faster Training
Parallel training
When word aligning, using mgiza12 with multiple threads significantly speed up word alignment.
MGIZA To use MGIZA with multiple threads in the Moses training script, add these arguments:
.../train-model.perl -mgiza -mgiza-cpus 8 ....
To enable it in the EMS, add this to the [TRAINING] section
[TRAINING]
training-options = "-mgiza -mgiza-cpus 8"
snt2cooc When running GIZA++ or MGIZA, the first stage involves running a program
called
snt2cooc
This requires approximately 6GB+ for typical Europarl-size corpora (1.8 million sentences). For
users without this amount of memory on their computers, an alternative version is included in
MGIZA:
snt2cooc.pl
To use this script, you must copy 2 files to the same place where snt2cooc is run:
snt2cooc.pl
snt2coocrmp
12
http://sourceforge.net/projects/mgizapp/
3.4.
Optimizing Moses
95
Add this argument when running the Moses training script:
.../train-model.perl -snt2cooc snt2cooc.pl
Parallel Extraction
Once word alignment is completed, the phrase table is created from the aligned parallel corpus.
There are 2 main ways to speed up this part of the training process.
Firstly, the training corpus and alignment can be split and phrase pairs from each part can be
extracted simultaneously. This can be done by simply using the argument -cores, e.g.,
.../train-model.perl -cores 4
Secondly, the Unix sort command is often executed during training. It is essential to optimize
this command to make use of the available disk and CPU. For example, recent versions of sort
can take the following arguments
sort
-S 10G --batch-size 253 --compress-program gzip --parallel 5
The Moses training script names these arguments
.../train-model.perl -sort-buffer-size 10G -sort-batch-size 253 \
-sort-compress gzip -sort-parallel 5
You should set these arguments. However, DO NOT just blindly copy the above settings, they
must be tuned to the particular computer you are running on. The most important issues are:
1. you must make sure the version of sort on your machine supports the arguments you
specify, otherwise the script will crash. The --parallel, --compress-program, and --batch-size
arguments have only recently been added to the sort command.
2. make sure you have enough memory when setting -sort-buffer-size. In particular, you should take into account other programs running on the computer. Also, two
or three simultaneous sort program will run (one to sort the extract file, one to sort
extract.inv, one to sort extract.o). If there is not enough memory because you’ve set
sort-buffer-size too high, your entire computer will likely crash.
96
3. Tutorials
3. the maximum number for the --batch-size argument is OS-dependent. For example, it
is 1024 on Linux, 253 on old Mac OSX, 2557 on new OSX.
4. on Mac OSX, using --compress-program can occasionally result in the following timeout
errors.
gsort: couldn’t create process for gzip -d: Operation timed out
3.4.4
Training Summary
In summary, to maximize speed on a large server with many cores and up-to-date software,
.../train-model.perl -mgiza -mgiza-cpus 8 -cores 10 \
-parallel -sort-buffer-size 10G -sort-batch-size 253 \
-sort-compress gzip -sort-parallel 10
To run on a laptop with limited memory
.../train-model.perl -mgiza -mgiza-cpus 2 -snt2cooc snt2cooc.pl \
-parallel -sort-batch-size 253 -sort-compress gzip
In the EMS, for large servers, this can be done by adding:
[TRAINING]
script = $moses-script-dir/training/train-model.perl training-options = "-mgiza -mgiza-cpus 8 -cores 10 \ -parallel -sort-buffer-size 10G -sort-batch-size 253 \ -sort-compress gzip -sort-parallel 10" parallel = yes For servers with older OSes, and therefore older sort commands: [TRAINING] script =$moses-script-dir/training/train-model.perl
training-options = "-mgiza -mgiza-cpus 8 -cores 10 -parallel"
parallel = yes
3.4.
Optimizing Moses
97
For laptops with limited memory:
[TRAINING]
script = moses-script-dir/training/train-model.perl training-options = "-mgiza -mgiza-cpus 2 -snt2cooc snt2cooc.pl \ -parallel -sort-batch-size 253 -sort-compress gzip" parallel = yes 3.4.5 Language Model Convert your language model to binary format. This reduces loading time and provides more control. Building a KenLM binary file See the KenLM web site13 for the time-memory tradeoff presented by the KenLM data structures. Use bin/build_binary (found in the same directory as moses and moses_chart) to convert ARPA files to the binary format. You can preview memory consumption with: bin/build_binary file.arpa This preview includes only the language model’s memory usage, which is in addition to the phrase table etc. For speed, use the default probing data structure. bin/build_binary file.arpa file.binlm To save memory, change to the trie data structure bin/build_binary trie file.arpa file.binlm To further losslessly compress the trie ("chop" in the benchmarks), use -a 64 which will compress pointers to a depth of up to 64 bits. 13 http://kheafield.com/code/kenlm/benchmark/ 98 3. Tutorials bin/build_binary -a 64 trie file.arpa file.binlm Note that you can also make this parameter smaller which will go faster but use more memory. Quantization will make the trie smaller at the expense of accuracy. You can choose any number of bits from 2 to 25, for example 10: bin/build_binary -a 64 -q 10 trie file.arpa file.binlm Note that quantization can be used independently of -a. Loading on-demand By default, language models fully load into memory at the beginning. If you are short on memory, you can use on-demand language model loading. The language model must be converted to binary format in advance and should be placed on LOCAL DISK, preferably SSD. For KenLM, you should use the trie data structure, not the probing data structure. If the LM for binarized using IRSTLM, append .mm to the file name and change the ini file to reflect this. Eg. change [feature] IRSTLM .... path=file.lm to [feature] IRSTLM .... path=file.lm.mm If the LM was binarized using KenLM, add the argument lazyken=true. Eg. from [feature] KENLM .... to 3.4. Optimizing Moses 99 [feature] KENLM .... lazyken=true 3.4.6 Suffix array Suffix arrays store the entire parallel corpora and word alignment information in memory, instead of the phrase table. The parallel corpora and alignment file is often much smaller than the phrase table. For example, for the Europarl German-English (gzipped files): de = 94MB en = 84MB alignment = 57MB phrase-based = 2.0GB hierarchical = 16.0GB Therefore, it is more memory efficient to store the corpus in memory, rather than the entire phrase-table. This is usually structured as a suffix array to enable fast extraction of translations. Translations are extracted as needed, usually per input test set, or per input sentence. Moses support two different implementations of suffix arrays, one for phrase-based models14 , [[one for hierarchical models -> AdvancedFeatures#ntoc43 ]]. 3.4.7 Cube Pruning Cube pruning limits the number of hypotheses created for each stack (or chart cell in chart decoding). It is essential for chart decoding (otherwise decoding will take a VERY long time) and an option in phrase-based decoding. In the phrase-based decoder, add: [search-algorithm] 1 [cube-pruning-pop-limit] 500 There is a speed-quality tradeoff, lower pop limit means less work for the decoder, so faster decoding but less accurate translation. 14 http://www.statmt.org/moses/?n=Moses.AdvancedFeatures#ntoc35 100 3.4.8 3. Tutorials Minimizing memory during training TODO: MGIZA with reduced memory sntcoc 3.4.9 Minimizing memory during decoding TODO Compile-time options These options can be added to the bjam command line, trading generality for performance. You should do a full rebuild with -a when changing the values of most of these options. Don’t use factors? Add --max-factors=1 Tailor KenLM’s maximum order to only what you need. If your highest-order language model has order 5, add --kenlm-max-order=5 Turn debug symbols off for speed and a little more memory. debug-symbols=off But don’t expect support from the mailing list until you rerun with debug symbols on! Don’t care about debug messages? --notrace Download tcmalloc15 and see BUILD-INSTRUCTIONS.txt in Moses for installation instructions. bjam will automatically detect tcmalloc’s presence and link against it for multi-threaded builds. Install Boost and zlib static libraries. Then link statically: 15 https://code.google.com/p/gperftools 3.4. Optimizing Moses 101 --static This may mean you have to install Boost and zlib yourself. Running single-threaded? Add threading=single. Using hierarchical or string-to-tree models, but none with source syntax? --unlabelled-source 3.4.10 Phrase-table types Moses has multiple phrase table implementations. The one that suits you best depends on the model you’re using (phrase-based or hierarchical/syntax), and how much memory your server has. Here is a complete list of the types: Memory - this read in the phrase table into memory. For phrase-based model and chart decoding. Note that this is much faster than Binary and OnDisk phrase table format, but it uses a lot of RAM. Binary - a phrase table is converted into a ’database’. Only the translations which are required are loaded into memory. Therefore, requiring less memory, but potentially slower to run. For phrase-based model OnDisk - reimplementation of Binary for chart decoding. SuffixArray - stores the parallel training data and word alignment in memory, instead of the phrase table. Extraction is done on the fly. Also have a feature where you can add parallel data while the decoder is running (’Dynamic Suffix Array’). For Phrase-based models. See Levenberg et al., (2010)16 . ALSuffixArray - Suffix array for hierarchical models. See Lopez (2008)17 . FuzzyMatch - Implementation of Koehn and Senellart (2010)18 . Hiero - like SCFG, but translation rules are in standard Hiero-style format Compact - for phrase-based model. See Junczys-Dowmunt (2012)19 . Subsection last modified on March 21, 2014, at 11:17 AM 16 http://homepages.inf.ed.ac.uk/miles/papers/naacl10b.pdf http://www.cs.jhu.edu/~alopez/talks/mtm2008-lopez.pdf 18 http://homepages.inf.ed.ac.uk/pkoehn/publications/tm-smt-amta2010.pdf 19 http://ufal.mff.cuni.cz/pbml/98/art-junczys-dowmunt.pdf 17 102 3. Tutorials 3.5 Experiment Management System 3.5.1 Introduction The Experiment Management System (EMS), or Experiment.perl, for lack of a better name, makes it much easier to perform experiments with Moses. There are many steps in running an experiment: the preparation of training data, building language and translation models, tuning, testing, scoring and analysis of the results. For most of these steps, a different tool needs to be invoked, so this easily becomes very messy. Here a typical example: This graph was automatically generated by Experiment.perl. All that needed to be done was to specify one single configuration file that points to data files and settings for the experiment. In the graph, each step is a small box. For each step, Experiment.perl builds a script file that gets either submitted to the cluster or run on the same machine. Note that some steps are quite 3.5. Experiment Management System 103 involved, for instance tuning: On a cluster, the tuning script runs on the head node a submits jobs to the queue itself. Experiment.perl makes it easy to run multiple experimental runs with different settings or data resources. It automatically detects which steps do not have to be executed again but instead which results from an earlier run can be re-used. Experiment.perl also offers a web interface to the experimental runs for easy access and comparison of experimental results. The web interface also offers some basic analysis of results, such as comparing the n-gram matches between two different experimental runs: 3.5.2 Requirements In order to run properly, EMS will require: • A version of Moses along with SRILM, • The GraphViz toolkit20 , • The ImageMagick toolkit21 , and 20 21 http://www.graphviz.org/ http://www.imagemagick.org/script/index.php 104 3. Tutorials • The GhostView tool22 . 3.5.3 Quick Start Experiment.perl is extremely simple to use: • • • • • • • • Find experiment.perl in scripts/ems Get a sample configuration file from someplace (for instance scripts/ems/example/config.toy). Set up a working directory for your experiments for this task (mkdir does it). Edit the following path settings in config.toy – working-dir – data-dir – moses-script-dir – moses-src-dir – srilm-dir – decoder Run experiment.perl -config config.toy from your experiment working directory. Marvel at the graphical plan of action. Run experiment.perl -config config.toy -exec. Check the results of your experiment (in evaluation/report.1) Let us take a closer look at what just happened. The configuration file config.toy consists of several sections. For instance there is a section for each language model corpus to be used. In our toy example, this section contains the following: [LM:toy] ### raw corpus (untokenized) # raw-corpus =toy-data/nc-5k.$output-extension The setting raw-corpus species the location of the corpus. The definition uses the variables$toy-data and $output-extension, which are also settings defined elsewhere in the configuration file. These variables are resolved, leading to the file path ems/examples/data/nc-5k.en in your Moses scripts directory. The authoritative definition of the steps and their interaction is in the file experiment.meta (in the same directory as experiment.perl: scripts/ems). The logic of experiment.meta is that it wants to create a report at the end. To generate the report it needs to evaluation scores, to get these it needs decoding output, to get these it needs to run the decoder, to be able to run the decoder it needs a trained model, to train a model it needs data. This process of defining the agenda of steps to be executed is very similar to the Make utility in Unix. We can find the following step definitions for the language model module in experiment.meta: 22 http://www.gnu.org/software/gv/ 3.5. Experiment Management System 105 get-corpus in: get-corpus-script out: raw-corpus default-name: lm/txt template: IN > OUT tokenize in: raw-corpus out: tokenized-corpus default-name: lm/tok pass-unless: output-tokenizer template:$output-tokenizer < IN > OUT
parallelizable: yes
The tokenization step tokenize requires raw-corpus as input. In our case, we specified the
setting in the configuration file. We could have also specified an already tokenized corpus with
tokenized-corpus. This would allow us to skip the tokenization step. Or, to give another
example, we could have not specified raw-corpus, but rather specify a script that generates the
corpus with the setting get-corpus-script. This would have triggered the creation of the step
get-corpus.
The steps are linked with the definition of their input in and output out. Each step has also a
default name for the output (efault-name) and other settings.
The tokenization step has as default name lm/tok. Let us look at the directory lm to see which
files it contains:
% ls -tr lm/*
lm/toy.tok.1
lm/toy.truecased.1
lm/toy.lm.1
We find the output of the tokenization step in the file lm/toy.tok.1. The toy was added from
the name definition of the language model (see [LM:toy] in config.toy). The 1 was added,
because this is the first experimental run.
The directory steps contains the script that executes each step, its STDERR and STDOUT output, and meta-information. For instance:
% ls steps/1/LM_toy_tokenize.1* | cat
steps/1/LM_toy_tokenize.1
steps/1/LM_toy_tokenize.1.DONE
steps/1/LM_toy_tokenize.1.INFO
106
3. Tutorials
steps/1/LM_toy_tokenize.1.STDERR
steps/1/LM_toy_tokenize.1.STDERR.digest
steps/1/LM_toy_tokenize.1.STDOUT
The file steps/2/LM_toy_tokenize.2 is the script that is run to execute the step. The file with
the extension DONE is created when the step is finished - this communicates to the scheduler that
subsequent steps can be executed. The file with the extension INFO contains meta-information
- essential the settings and dependencies of the step. This file is checked to detect if a step can
be re-used in new experimental runs.
In case that the step crashed, we expect some indication of a fault in STDERR (for instance the
words core dumped or killed). This file is checked to see if the step was executed successfully,
so subsequent steps can be scheduled or the step can be re-used in new experiments. Since the
STDERR file may be very large (some steps create megabytes of such output), a digested version
is created in STDERR.digest. If the step was successful, it is empty. Otherwise it contains the
error pattern that triggered the failure detection.
Let us now take a closer look at re-use. If we run the experiment again but change some of the
settings, say, the order of the language model, then there is no need to re-run the tokenization.
Here is the definition of the language model training step in experiment.meta:
train
in: split-corpus
out: lm
default-name: lm/lm
ignore-if: rlm-training
rerun-on-change: lm-training order settings
template: $lm-training -order$order settings -text IN -lm OUT error: cannot execute binary file The mention of order in the list behind rerun-on-change informs experiment.perl that this step does need to be re-run, if the order of the language model changes. Since none of the settings in the chain of steps leading up to the training have been changed, the step can be re-used. Try changing the language model order (order = 5 in config.toy), run experiment.perl again (experiment.perl -config config.toy) in the working directory, and you will see the new language model in the directory lm: % ls -tr lm/* lm/toy.tok.1 lm/toy.truecased.1 lm/toy.lm.1 lm/toy.lm.2 3.5. Experiment Management System 3.5.4 107 More Examples The example directory contains some additional examples. These require the training and tuning data released for the Shared Translation Task for WMT 2010. Create a working directory, and change into it. Then execute the following steps: mkdir data cd data wget http://www.statmt.org/wmt10/training-parallel.tgz tar xzf training-parallel.tgz wget http://www.statmt.org/wmt10/dev.tgz tar xzf dev.tgz cd .. The examples using these corpora are • • • • config.basic - a basic phrase based model, config.factored - a factored phrase based model, config.hierarchical - a hierarchical phrase based model, and config.syntax - a target syntax model. In all these example configuration files, most corpora are commented out. This is done by adding the word IGNORE at the end of a corpus definition (also for the language models). This allows you to run a basic experiment with just the News Commentary corpus which finished relatively quickly. Remove the IGNORE to include more training data. You may run into memory and disk space problems when using some of the larger corpora (especially the news language model), depending on your computing infrastructure. If you decide to use multiple corpora for the language model, you may also want to try out interpolating the individual language models (instead of using them as separate feature functions). For this, you need to comment out the IGNORE next to the [INTERPOLATED-LM] section. You may also specify different language pairs by changing the input-extension, output-extension, and pair-extension settings. Finally, you can run all the experiments with the different given configuration files and the data variations in the same working directory. The experimental management system figures out automatically which processing steps do not need to repeated because they can be re-used from prior experimental runs. Phrase Model Phrase models are, compared to the following examples, the simplest models to be trained with Moses and the fastest models to run. You may prefer these models over the more sophisticated models whose added complexity may not justify the small (if any) gains. The example config.basic is similar to the toy example, except for a larger training and test corpora. Also, the tuning stage is not skipped. Thus, even with most of the corpora commented out, the entire experimental run will likely take a day, with most time taken up by word alignment (TRAINING_run-giza and TRAINING_run-giza-inverse) and tuning (TUNING_tune). 108 3. Tutorials Factored Phrase Model Factored models allow for additional annotation at the word level which may be exploited in various models. The example in config.factored uses part-of-speech tags on the English target side. Annotation with part-of-speech tags is done with MXPOST, which needs to be installed first. Please read the installation instructions23 . After this, you can run experiment.perl with the configuration file config.factored. If you compare the factored example config.factored with the phrase-based example config.basic, you will notice the definition of the factors used: ### factored training: specify here which factors used # if none specified, single factor training is assumed # (one translation step, surface to surface) # input-factors = word output-factors = word pos alignment-factors = "word -> word" translation-factors = "word -> word+pos" reordering-factors = "word -> word" #generation-factors = decoding-steps = "t0" the factor definition: ################################################################# # FACTOR DEFINITION [INPUT-FACTOR] # also used for output factors temp-dir =working-dir/training/factor
[OUTPUT-FACTOR:pos]
### script that generates this factor
#
mxpost = /home/pkoehn/bin/mxpost
factor-script = "$moses-script-dir/training/wrappers/make-factor-en-pos.mxpost.perl -mxpost$mxpost"
and the specification of a 7-gram language model over part of speech tags:
23
http://www.statmt.org/moses/?n=Moses.ExternalTools#mxpost
3.5.
Experiment Management System
109
[LM:nc=pos]
factors = "pos"
order = 7
settings = "-interpolate -unk"
raw-corpus = $wmt10-data/training/news-commentary10.$pair-extension.$output-extension This factored model using all the available corpora is identical to the Edinburgh submission to the WMT 2010 shared task for English-Spanish, Spanish-English, and English-German language pairs (the French language pairs also used the 109 corpus, the Czech language pairs did not use the POS language model, and German-English used additional pre-processing steps). Hierarchical model Hierarchical phrase models allow for rules with gaps. Since these are represented by nonterminals and such rules are best processed with a search algorithm that is similar to syntactic chart parsing, such models fall into the class of tree-based or grammar-based models. For more information, please check the Syntax Tutorial (Section 3.3). From the view of setting up hierarchical models with experiment.perl, very little has to be changed in comparison to the configuration file for phrase-based models: % diff config.basic config.hierarchical 33c33 < decoder =$moses-src-dir/bin/moses
--> decoder = $moses-src-dir/bin/moses_chart 36c36 < ttable-binarizer =$moses-src-dir/bin/processPhraseTable
--> #ttable-binarizer = $moses-src-dir/bin/processPhraseTable 39c39 < #ttable-binarizer = "$moses-src-dir/bin/CreateOnDiskPt 1 1 5 100 2"
--> ttable-binarizer = "$moses-src-dir/bin/CreateOnDiskPt 1 1 5 100 2" 280c280 < lexicalized-reordering = msd-bidirectional-fe --> #lexicalized-reordering = msd-bidirectional-fe 284c284 < #hierarchical-rule-set = true --> hierarchical-rule-set = true 413c413 110 3. Tutorials < decoder-settings = "-search-algorithm 1 -cube-pruning-pop-limit 5000 -s 5000" --> #decoder-settings = "" The changes are: a different decoder binary (by default compiled into bin/moses_chart) and ttable-binarizer are used. The decoder settings for phrasal cube pruning do not apply. Also, hierarchical models do not allow for lexicalized reordering (their rules fulfill the same purpose), and the setting for hierarchical rule sets has to be turned on. The use of hierarchical rules is indicated with the setting hierarchical-rule-set. Target syntax model Syntax models imply the use of linguistic annotation for the non-terminals of hierarchical models. This requires running a syntactic parser. In our example config.syntax, syntax is used only on the English target side. The syntactic constituents are labeled with Collins parser, which needs to be installed first. Please read the installation instructions24 . Compared to the hierarchical model, very little has to be changed in the configuration file: % diff config.hierarchical config.syntax 46a47,49 > # syntactic parsers > collins = /home/pkoehn/bin/COLLINS-PARSER > output-parser = "$moses-script-dir/training/wrappers/parse-en-collins.perl"
>
241c244
< #extract-settings = ""
--> extract-settings = "--MinHoleSource 1 --NonTermConsecSource"
The parser needs to be specified, and the extraction settings may be adjusted. And you are
3.5.5
Try a Few More Things
Stemmed Word Alignment
The factored translation model training makes it very easy to set up word alignment not based
on the surface form of words, but any other property of a word. One relatively popular method
is to use stemmed words for word alignment.
24
http://www.statmt.org/moses/?n=Moses.ExternalTools#collins
3.5.
Experiment Management System
111
There are two reasons for this: For one, for morphologically rich languages, stemming overcomes data sparsity problems. Secondly, GIZA++ may have difficulties with very large vocabulary sizes, and stemming reduces the number of unique words.
To set up stemmed word alignment in experiment.perl, you need to define a stem as a factor:
[OUTPUT-FACTOR:stem4]
factor-script = "$moses-script-dir/training/wrappers/make-factor-stem.perl 4" [INPUT-FACTOR:stem4] factor-script = "$moses-script-dir/training/wrappers/make-factor-stem.perl 4"
and indicate the use of this factor in the TRAINING section:
input-factors = word stem4
output-factors = word stem4
alignment-factors = "stem4 -> stem4"
translation-factors = "word -> word"
reordering-factors = "word -> word"
#generation-factors =
decoding-steps = "t0"
GIZA++ is one of the slowest steps in the training pipeline. Qin Gao implemented a multithreaded version of GIZA++, called MGIZA, which speeds up word alignment on multi-core
machines.
To use MGIZA, you will first need to install25 it.
To use it, you simply need to add some training options in the section TRAINING:
### general options
#
training-options = "-mgiza -mgiza-cpus 8"
25
http://www.statmt.org/moses/?n=Moses.ExternalTools#mgiza
112
3. Tutorials
Using Berkeley Aligner
The Berkeley Aligner is a alternative to GIZA++ for word alignment. You may (or may not) get
better results using this tool.
To use the Berkeley Aligner, you will first need to install26 it.
The example configuration file already has a section for the parameters for the tool. You need
to un-comment them and adjust berkeley-jar to your installation. You should comment out
alignment-symmetrization-method, since this is a GIZA++ setting.
### symmetrization method to obtain word alignments from giza output
# (commonly used: grow-diag-final-and)
#
#alignment-symmetrization-method = grow-diag-final-and
### use of berkeley aligner for word alignment
#
use-berkeley = true
alignment-symmetrization-method = berkeley
berkeley-train = $moses-script-dir/ems/support/berkeley-train.sh berkeley-process =$moses-script-dir/ems/support/berkeley-process.sh
berkeley-jar = /your/path/to/berkeleyaligner-2.1/berkeleyaligner.jar
berkeley-java-options = "-server -mx30000m -ea"
berkeley-training-options = "-Main.iters 5 5 -EMWordAligner.numThreads 8"
berkeley-posterior = 0.5
The Berkeley Aligner proceeds in two step: a training step to learn the alignment model from
the data and a processing step to find the best alignment for the training data. This step has the
parameter berkeley-posterior to adjust a bias towards more or less alignment points. You
can try different runs with different values for this parameter. Experiment.perl will not re-run
the training step, just the processing step.
Using Dyer’s Fast Align
Another alternative to GIZA++ is fast_align from Dyer et al.27 . It runs much faster, and may
even give better results, especially for language pairs without much large-scale reordering.
To use Fast Align, you will first need to install28 it.
The example configuration file already has a example setting for the tool, using the recommended defaults. Just remove the comment marker @#@ before the setting:
26
http://www.statmt.org/moses/?n=Moses.ExternalTools#berkeley
http://www.ark.cs.cmu.edu/cdyer/fast_valign.pdf
28
http://www.statmt.org/moses/?n=Moses.ExternalTools#fastalign
27
3.5.
Experiment Management System
113
### use of Chris Dyer’s fast align for word alignment
#
fast-align-settings = "-d -o -v"
Experiment.perl assumes that you copied the binary into the usual external bin dir (setting
external-bin-dir) where GIZA++ and other external binaries are located.
IRST Language Model
The provided examples use the SRI language model during decoding. When you want to
use the IRSTLM instead, an additional processing step is required: the language model has to
converted into a binary format.
This part of the LM section defines the use of IRSTLM:
### script to use for binary table format for irstlm
# (default: no binarization)
#
#lm-binarizer = $moses-src-dir/irstlm/bin/compile-lm ### script to create quantized language model format (irstlm) # (default: no quantization) # #lm-quantizer =$moses-src-dir/irstlm/bin/quantize-lm
If you un-comment lm-binarizer, IRSTLM will be used. If you comment out in addition
lm-quantizer, the language model will be compressed into a more compact representation.
Note that the values above assume that you installed the IRSTLM toolkit in the directory
$moses-src-dir/irstlm. Randomized Language Model Randomized language models allow a much more compact (but lossy) representation. Being able to use much larger corpora for the language model may be beneficial over the small chance of making mistakes. First of all, you need to install29 the RandLM toolkit. There are two different ways to train a randomized language model. One is to train it from scratch. The other way is to convert a SRI language model into randomized representation. 29 http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#randlm 114 3. Tutorials Training from scratch: Find the following section in the example configuration files and uncomment the rlm-training setting. Note that the section below assumes that you installed the randomized language model toolkit in the directory$moses-src-dir/randlm.
### tool to be used for training randomized language model from scratch
# (more commonly, a SRILM is trained)
#
rlm-training = "$moses-src-dir/randlm/bin/buildlm -falsepos 8 -values 8" Converting SRI language model: Find the following section in the example configuration files and un-comment the lm-randomizer setting. ### script to use for converting into randomized table format # (default: no randomization) # lm-randomizer = "$moses-src-dir/randlm/bin/buildlm -falsepos 8 -values 8"
You may want to try other values for falsepos and values. Please see the language model
You can also randomize a interpolated language model by specifying the lm-randomizer in the
section INTERPOLATED-LM.
Compound Splitting
Compounding languages, such as German, allow the creation of long words such as Neuwortgenerierung (new word generation). This results in a lot of unknown words in any text, so splitting
up these compounds is a common method when translating from such languages.
Moses offers a support tool that splits up words, if the geometric average of the frequency of
its parts is higher than the frequency of a word. The method requires a model (the frequency
statistics of words in a corpus), so there is a training and application step.
Such word splitting can be added to experiment.perl simply by specifying the splitter script in
the GENERAL section:
input-splitter = moses-script-dir/generic/compound-splitter.perl Splitting words on the output side is currently not supported. 30 http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#randlm 3.5. Experiment Management System 3.5.6 115 A Short Manual The basic lay of the land is: experiment.perl breaks up the training, tuning, and evaluating of a statistical machine translation system into a number of steps, which are then scheduled to run in parallel or sequence depending on their inter-dependencies and available resources. The possible steps are defined in the file experiment.meta. An experiment is defined by a configuration file. The main modules of running an experiment are: • • • • • • • • • • • CORPUS: preparing a parallel corpus, INPUT-FACTOR and OUTPUT-FACTOR: commands to create factors, TRAINING: training a translation model, LM: training a language model, INTERPOLATED-LM: interpolate language models, SPLITTER: training a word splitting model, RECASING: training a recaser, TRUECASING: training a truecaser, TUNING: running minumum error rate training to set component weights, TESTING: translating and scoring a test set, and REPORTING: compile all scores in one file. Experiment.Meta The actual steps, their dependencies and other salient information are to be found in the file experiment.meta. Think of experiment.meta as a "template" file. Here the parts of the step description for CORPUS:get-corpus and CORPUS:tokenize: get-corpus in: get-corpus-script out: raw-stem [...] tokenize in: raw-stem out: tokenized-stem [...] Each step takes some input (in) and provides some output (out). This also establishes the dependencies between the steps. The step tokenize requires the input raw-stem. This is provided by the step get-corpus. experiment.meta provides a generic template for steps and their interaction. For an actual experiment, a configuration file determines which steps need to be run. This configuration file is the one that is specified when invocing experiment.perl. It may contain for instance the following: 116 3. Tutorials [CORPUS:europarl] ### raw corpus files (untokenized, but sentence aligned) # raw-stem =europarl-v3/training/europarl-v3.fr-en
Here, the parallel corpus to be used is named europarl and it is provided in raw text format in the location $europarl-v3/training/europarl-v3.fr-en (the variable$europarl-v3
is defined elsewhere in the config file). The effect of this specification in the config file is that
the step get-corpus does not need to be run, since its output is given as a file. More on the
configuration file below in the next section.
Several types of information are specified in experiment.meta:
• in and out: Established dependencies between steps; input may also be provided by files
specified in the configuration.
• default-name: Name of the file in which the output of the step will be stored.
• template: Template for the command that is placed in the execution script for the step.
• template-if: Potential command for the execution script. Only used, if the first parameter exists.
• error: experiment.perl detects if a step failed by scanning STDERR for key words such
with this parameter.
• not-error: Declares default error key words as not indicating failures.
• pass-unless: Only if the given parameter is defined, this step is executed, otherwise the
step is passed (illustrated by a yellow box in the graph).
• ignore-unless: If the given parameter is defined, this step is not executed. This overrides
requirements of downstream steps.
• rerun-on-change: If similar experiments are run, the output of steps may be used, if
input and parameter settings are the same. This specifies a number of parameters whose
change disallows a re-use in different run.
• parallelizable: When running on the cluster, this step may be parallelized (only if
generic-parallelizer is set in the config file, the script can be found in $moses-script-dir/scripts/em • qsub-script: If running on a cluster, this step is run on the head node, and not submitted to the queue (because it submits jobs itself). Here now the full definition of the step CONFIG:tokenize tokenize in: raw-stem out: tokenized-stem default-name: corpus/tok pass-unless: input-tokenizer output-tokenizer template-if: input-tokenizer IN.$input-extension OUT.$input-extension template-if: output-tokenizer IN.$output-extension OUT.$output-extension parallelizable: yes 3.5. Experiment Management System 117 The step takes raw-stem and produces tokenized-stem. It is parallizable with the generic parallelizer. That output is stored in the file corpus/tok. Note that the actual file name also contains the corpus name, and the run number. Also, in this case, the parallel corpus is stored in two files, so file name may be something like corpus/europarl.tok.1.fr and corpus/europarl.tok.1.en. The step is only executed, if either input-tokenizer or output-tokenizer are specified. The templates indicate how the command lines in the execution script for the steps look like. Multiple Corpora, One Translation Model We may use multiple parallel corpora for training a translation model or multiple monolingual corpora for training a language model. Each of these have their own instances of the CORPUS and LM module. There may be also multiple test sets in TESTING). However, there is only one translation model and hence only one instance of the TRAINING module. The definitions in experiment.meta reflect the different nature of these modules. For instance CORPUS is flagged as multiple, while TRAINING is flagged as single. When defining settings for the different modules, the singular module TRAINING has only one section, while this one general section and specific LM sections for each training corpus. In the specific section, the corpus is named, e.g. LM:europarl. As you may imagine, the tracking of dependencies between steps of different types of modules and the consolidation of corpus-specific instances of modules is a bit complex. But most of that is hidden from the user of the Experimental Management System. When looking up the parameter settings for a step, first the set-specific section (LM:europarl) is consulted. If there is no definition, then the module definition (LM) and finally the general definition (in section GENERAL) is consulted. In other words, local settings override global settings. Defining Settings The configuration file for experimental runs is a collection of parameter settings, one per line with empty lines and comment lines for better readability, organized in sections for each of the modules. The syntax of setting definition is setting = value (note: spaces around the equal sign). If the value contains spaces, it must be placed into quotes (setting = "the value"), except when a vector of values is implied (only used when defining list of factors: output-factor = word pos. Comments are indicated by a hash (#). The start of sections is indicated by the section name in square brackets ([TRAINING] or [CORPUS:europarl]). If the word IGNORE is appended to a section definition, then the entire section is ignored. Settings can be used as variables to define other settings: 118 3. Tutorials working-dir = /home/pkoehn/experiment wmt10-data =$working-dir/data
Variable names may be placed in curly brackets for clearer separation:
wmt10-data = ${working-dir}/data Such variable references may also reach other modules: [RECASING] tokenized =$LM:europarl:tokenized-corpus
Finally, reference can be made to settings that are not defined in the configuration file, but are
the product of the defined sequence of steps.
Say, in the above example, tokenized-corpus is not defined in the section LM:europarl, but instead raw-corpus. Then, the tokenized corpus is produced by the normal processing pipeline.
Such an intermediate file can be used elsewhere:
[RECASING]
tokenized = [LM:europarl:tokenized-corpus]
Some error checking is done on the validity of the values. All values that seem to be file paths
trigger the existence check for such files. A file with the prefix of the value must exist.
There are a lot of settings reflecting the many steps, and explaining these would require explaining the entire training, tuning, and testing pipeline. Please find the required documentation for
step elsewhere around here. Every effort has been made to include verbose descriptions in the
example configuration files, which should be taken as starting point.
Working with Experiment.Perl
You have to define an experiment in a configuration file and the Experiment Management
System figures out which steps need to be run and schedules them either as jobs on a cluster or
runs them serially on a single machine.
Other options:
3.5.
Experiment Management System
119
• -no-graph: Supresses the display of the graph.
• -continue RUN: Continues the experiment RUN, which crashed earlier. Make sure that
crashed step and its output is deleted (see more below).
• -delete-crashed RUN: Delete all step files and their output files for steps that have crashed
in a particular RUN.
• -delete-run RUN: Delete all step files and their output files for steps for a given RUN,
unless these steps are used by other runs.
• -delete-version RUN: Same as above.
• -max-active: Specifies the number of steps that can be run in parallel when running on
a single machine (default: 2, not used when run on cluster).
• -sleep: Sets the number of seconds to be waited in the scheduler before the completion
of tasks is checked (default: 2).
• -ignore-time: Changes the re-use behavior. By default files cannot be re-used when
their time stamp changed (typically a tool such as the tokenizer which was changed, thus
requiring re-running all tokenization steps in new experiments). With this switch, files
with changed time stamp can be re-used.
• -meta: Allows the specification of a custom experiment.meta file, instead of using the
one in the same directory as the experiment.perl script.
• -final-step STEP: Do not run a complete experiment, but finish at the specified STEP.
• -final-out OUT: Do not run a complete experiment, but finish when the specified output
file OUT is created. These are the output file specifiers as used in experiment.meta.
• -cluster: Indicates that the current machine is a cluster head node. Step files are submitted as jobs to the cluster.
• -multicore: Indicates that the current machine is a multi-core machine. This allows for
additional parallelization with the generic parallelizer setting.
The script may automatically detect if it is run on a compute cluster or a multi-core machine, if
this is specified in the file experiment.machines, for instance:
cluster: townhill seville
multicore-8: tyr thor
multicore-16: loki
defines the machines townhill and seville as GridEngine cluster machines, tyr and thor as
8-core machines and loki as 16-core machines.
Typically, experiments are started with the command:
experiment.perl -config my-config -exec
Since experiments run for a long time, you may want to run this in the background and also
set a nicer priority:
120
3. Tutorials
nice nohup -config my-config -exec >& OUT.[RUN] &
This keeps also a report (STDERR and STDOUT) on the execution in a file named, say, OUT.1,
with the number corresponding to the run number.
The meta-information for the run is stored in the directory steps. Each run has a sub directory
with its number (steps/1, steps/2, etc.). The sub directory steps/0 contains step specification
when Experiment.perl is called without the -exec switch.
The sub directories for each run contain the step definitions, as well as their meta-information
and output. The sub directories also contain a copy of the configuration file (e.g. steps/1/config.1),
the agenda graph (e.g. steps/1/graph.1.{dot,ps,png}), a file containing all expanded parameter settings (e.g. steps/1/parameter.1), and an empty file that is touched every minute
as long as the experiment is still running (e.g. steps/1/running.1).
Continuing Crashed Experiments
Steps may crash. No, steps will crash, be it because faulty settings, faulty tools, problems with
the computing resources, willful interruption of an experiment, or an act of God.
The first thing to continue a crashed experiment is to detect the crashed step. This is shown
either by the red node in the displayed graph or reported on the command line in the last
lines before crashing; though this may not be pretty obvious, if parallel steps kept running
after that. However, the automatic error detection is not perfect and a step may have failed
upstream without detection causes failure further down the road.
You should have a understanding of what each step does. Then, by looking at its STDERR
and STDOUT file, and the output files it should have produced, you can track down what went
wrong.
Fix the problem, and delete all files associated with the failed step (e.g., rm steps/13/TUNING_tune.13*,
rm -r tuning/tmp.1). To find what has been produced by the crashed step, you may need to
consult where the output of this step is placed, by looking at experiment.meta.
You can automatically delete all crashed steps and their output files with
experiment.perl -delete-crashed 13 -exec
After removing the failed step and ensuring that the cause of the crash has been addressed,
you can continue a crashed experimental run (e.g., run number 13) with:
experiment.perl -continue 13 -exec
3.5.
Experiment Management System
121
You may want to check what will be run by excluding the -exec command at first. The graph
indicates which steps will be re-used from the original crashed run.
If the mistake was a parameter setting, you can change that setting in the stored configuration
file (e.g., steps/1/config.1). Take care, however, to delete all steps (and their subsequent
steps) that would have been run differently with that setting.
If an experimental run crashed early, or you do not want to repeat it, it may be easier to delete
the entire step directory (rm -r steps/13). Only do this with the latest experimental run (e.g.,
not when there is already a run 14), otherwise it may mess up the re-use of results.
You may also delete all output associated with a run with the command rm -r */*.13*. However this requires some care, so you may want to check first what you are deleting (ls */*.13).
Running a Partial Experiment
By default, experiment.perl will run a full experiment: model building, tuning and testing. You
may only want to run parts of the pipeline, for instance building a model, but not tuning and
testing. You can do this by specifying either a final step or a final outcome.
If you want to terminate at a specific step
experiment.perl -config my-config -final-step step-name -exec
where step-name is for instance TRAINING:create-config, LM:my-corpus:train, or TUNING:tune.
If you want to terminate once a particular output file is generated:
experiment.perl -config my-config -final-out out -exec
Examples for out are TRAINING:config, LM:my-corpus:lm, or TUNING:weight-config. In fact,
these three examples are identical to the three examples above, it is just another way to specify
the final point of the pipeline.
Technically, this works by not using REPORTING:report as the end point of the pipeline, but the
specified step.
Removing a Run
If you want to remove all the step files and output files associated with a particular run, you
can do this with, for instance:
experiment.perl -delete-run
13 -exec
122
3. Tutorials
If you run this without -exec you will see a list of files that would be deleted (but no files are
actually deleted).
Steps that are used in other runs, and the output files that they produced are kept. Also, the
step directory (e.g., steps/13 is not removed. You may remove this by hand, if there are no
step files left.
Running on a Cluster
Experiment.perl works with Sun GridEngine clusters. The script needs to be run on the head
node and jobs are scheduled on the nodes.
There are two ways to tell experiment.perl that the current machine is a cluster computer. One
is by using the switch -cluster, or by adding the machine name into experiment.machines.
The configuration file has a section that allows for the setting of cluster-specific settings. The
setting jobs is used to specify into how many jobs to split the decoding during tuning and
testing. For more details on this, please see moses-parallel.pl.
All other settings specify switches that are passed along with each submission of a job via qsub:
•
•
•
•
qsub-memory: number of memory slots (-pe memory NUMBER),
qsub-hours: number of hours reserved for each job (-l h_rt=NUMBER:0:0),
qsub-project: name if the project for user accounting (-P PROJECT), and
qsub-settings: any other setting that is passed along verbatim.
Note that the general settings can be overriden in each module definition - you may want to
have different settings for different steps.
If the setting generic-parallelizer is set (most often it is set to to the ems support script
$moses-script-dir/ems/support/generic-parallelizer.perl), then a number of additional steps are parallelized. For instance, tokenization is performed by breaking up the corpus into as many parts as specified with jobs, jobs to process the parts are submitted in parallel to the cluster, and their output pieced together upon completion. Be aware that there are many different ways to configure a GridEngine cluster. Not all the options described here may be available, and it my not work out of the box, due to your specific installation. Running on a Multi-core Machine Using a multi-core machine means first of all that more steps can be scheduled in parallel. There is also a generic parallelizer (generic-multicore-parallelizer.perl) that plays the same role as the generic parallelizer for clusters. However, decoding is not broken up into several pieces. It is more sensible to use multithreading in the decoder31 . 31 http://www.statmt.org/moses/?n=Moses.AdvancedFeatures#multi-threaded 3.5. Experiment Management System 123 Web Interface The introduction included some screen shots of the web interface to the Experimental Management System. You will need to have a running web server on a machine (LAMPP on Linux or MAMP on Mac does the trick) that has access to the file system where your working directory is stored. Copy or link the web directory (in scripts/ems) on a web server. Make sure the web server user has the right write permissions on the web interface directory. To add your experiments to this interface, add a line to the file /your/web/interface/dir/setup. The format of the file is explained in the file. 3.5.7 Analysis You can include additional analysis for an experimental run into the web interface by specifying the setting analysis in its configuration file. analysis =$moses-script-dir/ems/support/analysis.perl
This currently reports n-gram precision and recall statistics and color-coded n-gram correctness
markup for the output sentences, as in
The output is color-highlighted according to n-gram matches with the reference translation.
The following colors are used:
•
•
•
•
•
grey: word not in reference,
light blue: word part of 1-gram match,
blue: word part of 2-gram match,
dark blue: word part of 3-gram match, and
very dark blue: word part of 4-gram match.
Segmentation
The setting analyze-coverage include a coverage analysis: which words and phrases in the
input occur in the training data or the translation table? This is reported in color coding and
in a yellow report box when moving the mouse of the word or the phrase. Also, summary
statistics for how many words occur how often are given, and a report on unknown or rare
words is generated.
124
3. Tutorials
Coverage Analysis
The setting analyze-coverage include a coverage analysis: which words and phrases in the
input occur in the training data or the translation table? This is reported in color coding and
in a yellow report box when moving the mouse of the word or the phrase. Also, summary
statistics for how many words occur how often are given, and a report on unknown or rare
words is generated.
Bilingual Concordancer
To more closely inspect where input words and phrases occur in the training corpus, the analysis tool includes a bilingual concordancer. You turn it on by adding this line to the training
biconcor = moses-bin-dir/biconcor During training, a suffix array of the corpus is built in the model directory. The analysis web interface accesses these binary files to quickly scan for occurrences of source words and phrases in the training corpus. For this to work, you need to include the biconcor binary in the web root directory. When you click on a word or phrase, the web page is augemented with a section that shows all (or frequent word, a sample of all) occurences of that phrase in the corpus, and how it was aligned: Source occurrences (with context) are shown on the left half, the aligned target on the right. In the main part, occurrences are grouped by different translations -– also shown bold in context. Unaligned boundary words are shown in blue. The extraction heuristic extracts additional rules for these cases, but these are not listed here for clarity. 3.5. Experiment Management System 125 At the end, source occurrences for which no rules could be extracted are shown. This may happen because the source words are not aligned to any target words. In this case, the tool shows alignments of the previous word (purple) and following word(olive), as well as some neighboring unaligned words (again, in blue). Another reason for failure to extract rules are misalignments, when the source phrase maps to a target span which contains words that also align to outside source words (violation of the coherence contraint). These misaligned words (in source and target) are shown in red. Precision by coverage To investigate further, if the correctness of the translation of input words depends on frequency in the corpus (and what the distribution of word frequency is), a report for precision by coverage can be turned on with the following settings: report-precision-by-coverage = yes precision-by-coverage-factor = pos precision-by-coverage-base =working-dir/evaluation/test.analysis.5
Only the first setting report-precision-by-coverage is needed for the report. The second
setting precision-by-coverage-factor provides an additional breakdown for a specific input
factor (in the example, the part-of-speech factor named pos). More on the precision-by-coverage-base
below.
When clicking on "precision of input by coverage" on the main page, a precision by coverage
graph is shown:
The log-coverage class is on the x-axis (-1 meaning unknown, 0 singletons, 1 words that occur
twice, 2 words that occur 3-4 times, 3 words that occur 5-8 times, and so on). The scale of boxes
for each class is determined by the ratio of words in the class in the test set. The precision of
translations of words in a class is shown on the y-axis.
Translation of precision of input words cannot be determined in a clear cut word. Our determination relies on phrase alignment of the decoder, word alignment within phrases, and accounting for multiple occurrences of transled words in output and reference translations. Not that
126
3. Tutorials
the precision metric does not penalize for dropping words, so this is shown in a second graph
(in blue), below the precision graph.
If you click on the graph, you will see the graph in tabular form. Following additional links
allows you to see breakdowns for the actual words, and even find the sentences in which they
occur.
Finally, the precision-by-coverage-base setting. For comparison purposes, it may be useful
to base the coverage statistics on the corpus of a previous run. For instance, if you add training
data, does the translation quality of the words increase? Well, a word that occured 3 times in
the small corpus, may now occur 10 times in the big corpus, hence the word is placed in a
different class. To maintain the original classification of words into the log-coverage classes,
you may use this setting to point to an earlier run.
4
User Guide
4.1
Support Tools
4.1.1
Overview
Scripts are in the scripts subdirectory in the source release in the Git repository.
The following basic tools are described elsewhere:
•
•
•
•
4.1.2
Moses decoder (Section 3.1)
Training script train-model.perl (Section 5.3)
Corpus preparation clean-corpus-n.perl (Section 5.2)
Minimum error rate training (tuning) mert-moses.pl (Section 5.14)
Converting Pharaoh configuration files to Moses configuration files
Moses is a successor the the Pharaoh decoder, so you can use the same models that work for
Pharaoh and use them with Moses. The following script makes the necessary changes to the
configuration file:
exodus.perl < pharaoh.ini > moses.ini
4.1.3
Moses decoder in parallel
Since decoding large amounts of text takes a long time, you may want to split up the text into
blocks of a few hundred sentences (or less), and distribute the task across a Sun GridEngine
cluster. This is supported by the script moses-parallel.pl, which is run as follows:
moses-parallel.pl -decoder decoder -config cfgfile -i input -jobs N [options]
127
128
4. User Guide
Use absolute paths for your parameters (decoder, configuration file, models, etc.).
•
•
•
•
•
4.1.4
decoder is the file location of the binary of Moses used for decoding
cfgfile is the configuration fileofthe decoder
input is the file to translate
N is the number of processors you require
options are used to overwrite parameters provided in cfgfile
Among them, overwrite the following two parameters for nbest generation (NOTE: they
differ from standard Moses)
– -n-best-file output file for nbest list
– -n-best-size size of nbest list
Filtering phrase tables for Moses
Phrase tables easily get too big, but for the translation of a specific set of text only a fraction of
the table is needed. So, you may want to filter the translation table, and this is possible with
the script:
filter-model-given-input.pl filter-dir config input-file
This creates a filtered translation table with new configuration file in the directory filter-dir
from the model specified with the configuration file config (typically named moses.ini), given
the (tokenized) input from the file input-file.
In the advanced feature section, you find the additional option of binarizing translation and
reordering table, which allows these models to be kept on disk and queried by the decoder. If
you want to both filter and binarize these tables, you can use the script:
filter-model-given-input.pl filter-dir config input-file -Binarizer binarizer
The additional binarizer option points to the appropriate version of processPhraseTable.
4.1.5
Reducing and Extending the Number of Factors
Instead of the two following scripts, this one does both at the same time, and is better suited
for our directory structure and factor naming conventions:
reduce_combine.pl \
czeng05.cs \
0,2 pos lcstem4 \
4.1.
4.1.6
Support Tools
129
Scoring translations with BLEU
A simple BLEU scoring tool is the script multi-bleu.perl:
multi-bleu.perl reference < mt-output
Reference file and system output have to be sentence-aligned (line X in the reference file corresponds to line X in the system output). If multiple reference translation exist, these have to
be stored in seperate files and named reference0, reference1, reference2, etc. All the texts
need to be tokenized.
A popular script to score translations with BLEU is the NIST mteval script1 . It requires that text
is wrapped into a SGML format. This format is used for instance by the NIST evaluation2 and
the WMT Shared Task evaluations3 . See the latter for more details on using this script.
4.1.7
Missing and Extra N-Grams
Missing n-grams are those that all reference translations wanted but MT system did not produce. Extra n-grams are those that the MT system produced but none of the references approved.
missing_and_extra_ngrams.pl hypothesis reference1 reference2 ...
4.1.8
Making a Full Local Clone of Moses Model + ini File
Assume you have a moses.ini file already and want to run an experiment with it. Some
months from now, you might still want to know what exactly did the model (incl. all the
tables) look like, but people tend to move files around or just delete them.
To solve this problem, create a blank directory, go in there and run:
clone_moses_model.pl ../path/to/moses.ini
close_moses_model.pl will make a copy of the moses.ini file and local symlinks (and if possible also hardlinks, in case someone deleted the original file) to all the tables and language
models needed.
It will be now safe to run moses locally in the fresh directory.
1
http://www.nist.gov/speech/tests/mt/2008/scoring.html
http://www.nist.gov/speech/tests/mt/2009/
3
2
130
4.1.9
4. User Guide
Absolutizing Paths in moses.ini
Run:
absolutize_moses_model.pl
../path/to/moses.ini > moses.abs.ini
to build an ini file where all paths to model parts are absolute. (Also checks the existence of the
files.)
4.1.10
The script
analyse_moses_model.pl moses.ini
Prints basic statistics about all components mentioned in the moses.ini. This can be useful to
set the order of mapping steps to avoid explosion of translation options or just to check that the
model components are as big/detailed as we expect.
Sample output lists information about a model with 2 translation and 1 generation step. The
three language models over three factors used and their n-gram counts (after discounting) are
listed, too.
Translation 0 -> 1 (/fullpathto/phrase-table.0-1.gz):
743193
phrases total
1.20 phrases per source phrase
Translation 1 -> 2 (/fullpathto/phrase-table.1-2.gz):
558046
phrases total
2.75 phrases per source phrase
Generation 1,2 -> 0 (/fullpathto/generation.1,2-0.gz):
1.04 outputs per source token
Language model over 0 (/fullpathto/lm.1.lm):
1
2
3
49469 245583 27497
Language model over 1 (/fullpathto/lm.2.lm):
1
2
3
25459 199852 32605
Language model over 2 (/fullpathto/lm.3.lm):
1
2
3
4
5
6
7
709
20946
39885
45753
27964
12962
7524
4.1.
Support Tools
4.1.11
131
Recaser
Often, we train machine translation systems on lowercased data. If we want to present the
output to a user, we need to re-case (or re-capitalize) the output. Moses provides a simple
tool to recase data, which essentially runs Moses without reordering, using a word-to-word
translation model and a cased language model.
The recaser requires a model (i.e., the word mapping model and language model mentioned
above), which is trained with the command:
train-recaser.perl --dir MODEL --corpus CASED [--ngram-count NGRAM] [--train-script TRAIN]
The script expects a cased (but tokenized) training corpus in the file CASED, and creates a recasing model in the directory MODEL. Since the script needs to run SRILM’s ngram-count and the
Moses train-model.perl, these need to be specified, otherwise the hard-coded defaults are
used.
To recase output from the Moses decoder, you run the command
recase.perl --in IN --model MODEL/moses.ini --moses MOSES [--lang LANGUAGE] [--headline SGML] > OUT
The input is in file IN, the output in file OUT. You also need to specify a recasing model MODEL.
Since headlines are capitalized different from regular text, you may want to provide an SGML file
that contains information about headline. This file uses the NIST format, and may be identical
to source test sets provided by the NIST or other evluation campaigns. A language LANGUAGE
may also be specified, but only English (en) is currently supported.
4.1.12
Truecaser
Instead of lowercasing all training and test data, we may also want to keep words in their natural case, and only change the words at the beginning of their sentence to their most frequent
form. This is what we mean by truecasing. Again, this requires first the training of a truecasing
model, which is a list of words and the frequency of their different forms.
train-truecaser.perl --model MODEL --corpus CASED
The model is trained from the cased (but tokenized) training corpus CASED and stored in the
file MODEL.
Input to the decoder has to be truecased with the command
132
4. User Guide
truecase.perl --model MODEL < IN > OUT
Output from the decoder has to be restored into regular case. This simply uppercases words at
the beginning of sentences:
detruecase.perl < in > out [--headline SGML]
An SGML file with headline information may be provided, as done with the recaser.
4.1.13
Searchgraph to DOT
This small tool converts Moses searchgraph (-output-search-graph FILE option) to dot format. The dot format can be rendered using the graphviz4 tool dot.
moses ... --output-search-graph temp.graph -s 3
# we suggest to use a very limited stack size, -s 3
sg2dot.perl [--organize-to-stacks] < temp.graph > temp.dot
dot -Tps temp.dot > temp.ps
Using --organize-to-stacks makes nodes in the same stack appear in the same column (this
slows down the rendering, off by default).
Caution: the input must contain the searchgraph of one sentence only.
4.1.14
Threshold Pruning of Phrase Table
The phrase table trained by Moses contains by default all phrase pairs encountered in the parallel training corpus. This often includes 100,000 different translations for the word "the" or the
comma ",". These may clog up various processing steps down the road, so it is helpful to prune
the phrase table to the reasonable choices.
Threshold pruning is currently implemented at two different stages: You may filter the entire
phrase table file, or use threshold pruning as an additional filtering criterion when filtering the
phrase table for a given test set. In either case, phrase pairs are thrown out when their phrase
translation probability p(e|f) falls below a specified threshold. A safe number for this threshold
may be 0.0001, in the sense that it hardly changes any phrase translation while ridding the table
of a lot of junk.
4
http://www.graphviz.org/
4.1.
Support Tools
133
Pruning the full phrase table file
The script scripts/training/threshold-filter.perl operates on any phrase table file:
cat PHRASE_TABLE | \
threshold-filter.perl 0.0001 > PHRASE_TABLE.reduced
If the phrase table is zipped, then:
zcat PHRASE_TABLE.gz | \
threshold-filter.perl 0.0001 | \
gzip - > PHRASE_TABLE.reduced.gz
While this often does not remove much of the phrase table (which contains to a large part singleton phrase pairs with p(e|f)=1), it may nevertheless be helpful to also reduce the reordering
model. This can be done with a second script:
cat REORDERING_TABLE | \
remove-orphan-phrase-pairs-from-reordering-table.perl PHRASE_TABLE \
> REORDERING_TABLE.pruned
Again, this also works for zipped files:
zcat REORDERING_TABLE.gz | \
remove-orphan-phrase-pairs-from-reordering-table.perl PHRASE_TABLE | \
gzip - > REORDERING_TABLE.pruned.gz
Pruning during test/tuning set filtering
In the typical experimental setup, the phrase table is filtered for a tuning or test set using the
script. During this process, we can also remove low-probability phrase pairs. This can be done
simply by adding the switch -MinScore, which takes a specification of the following form:
filter-model-given-input.pl [...] \
-MinScore FIELD1:THRESHOLD2[,FIELD2:THRESHOLD2[,FIELD3:THRESHOLD3]]
134
4. User Guide
where FIELDn is the position of the score (typically 2 for the direct phrase probability p(e|f), or
0 for the indirect phrase probability p(f|e)) and THRESHOLD the maximum probability allowed.
4.2
External Tools
A very active community is engaged in statistical machine translation research, which has
produced a number of tools that may be useful for training a Moses system. Also, the more
linguistically motivated models (factored model, syntax model) require tools to the linguistic
annotation of corpora.
In this section, we list some useful tools. If you know (or are the developer of) anything we
MT tools, refer to the following pages:
• List of Free/Open-source MT Tools5 , maintained by Mikel Forcada.
• TAUS Tracker6 , a comprehensive list of Translation and Language Technology tools maintained by TAUS.
4.2.1
Word Alignment Tools
Berkeley Word Aligner
The BerkeleyAligner7 (available at Sourceforge8 ) is a word alignment software package that
implements recent innovations in unsupervised word alignment. It is implemented in Java
and distributed in compiled format.
Installation:
mkdir /my/installation/dir
cd /my/installation/dir
tar xzf berkeleyaligner_unsupervised-2.1.tar.gz
Test:
cd berkeleyaligner
chmod +x align
./align example.conf
5
http://www.fosmt.org
http://www.taustracker.com/
7
http://nlp.cs.berkeley.edu/Main.html#WordAligner
8
6
4.2.
External Tools
135
MGIZA9 was developed by Qin Gao. It is an implementation of the popular GIZA++ word
alignment toolkit to run multi-threaded on multi-core machines. Check the web site for more
Installation:
svn checkout http://svn.code.sf.net/p/mgizapp/code/trunk mgizapp-code
cd mgizapp-code/mgizapp/
cmake .
make
make install
Compiling MGIZA requires the Boost library. If your Boost library are in non-system directory,
use the script
manual-compile/compile.sh
to compile MGIZA.
The MGIZA binary and the script merge_alignment.py need to be copied in you binary directory that Moses will look up for word alignment tools. This is the exact command I use to copy
MGIZA to it final destination:
export BINDIR=~/workspace/bin/training-tools
cp bin/* BINDIR/mgizapp cp scripts/merge_alignment.pyBINDIR
MGIZA works with the training script train-model.perl. You indicate its use (opposed to regular GIZA++) with the switch -mgiza. The switch -mgiza-cpus NUMBER allows you to specify
the number of CPUs.
Dyer et al.’s Fast Align
The Fast Align10 is a comparable fast unsupervised word aligner that nevertheless gives comparable results to GIZA++. It’s details are described in a NAACL 2013 paper11
Installation:
9
http://geek.kyloo.net/software/doku.php
11
http://www.ark.cs.cmu.edu/cdyer/fast_valign.pdf
10
136
4. User Guide
mkdir /my/installation/dir
cd /my/installation/dir
git clone https://github.com/clab/fast_align.git
cd fast_align
make
4.2.2
Evaluation Metrics
Translation Error Rate (TER)
Translation Error Rate12 is an error metric for machine translation that measures the number of
edits required to change a system output into one of the references. It is implemented in Java.
Installation:
mkdir /my/installation/dir
cd /my/installation/dir
wget http://www.cs.umd.edu/~snover/tercom/tercom-0.7.25.tgz
tar xzf tercom-0.7.25.tgz
METEOR
METEOR13 is a metric that includes stemmed and synonym matches when measuring the similarity between system output and human reference translations.
Installation:
mkdir /my/installation/dir
cd /my/installation/dir
wget http://www.cs.cmu.edu/~alavie/METEOR/install-meteor-1.0.sh
sh install-meteor-1.0.sh
12
13
http://www.cs.umd.edu/~snover/tercom/
http://www.cs.cmu.edu/~alavie/METEOR/
4.2.
4.2.3
External Tools
137
Part-of-Speech Taggers
MXPOST (English)
MXPOST was developed by Adwait Ratnaparkhi as part of his PhD thesis. It is a Java implementation of a maximum entropy model and distributed as compiled code. It can be trained
for any language pair for with annotated POS data exists.
Installation:
mkdir /your/installation/dir
cd /your/installation/dir
tar xzf jmx.tar.gz
echo ’#!/bin/ksh’ > mxpost
echo ’export CLASSPATH=/your/installation/dir/mxpost.jar’ >> mxpost
echo ’java -mx30m tagger.TestTagger /your/installation/dir/tagger.project’ >> mxpost
Test:
echo ’This is a test .’ | ./mxpost
The script script/training/wrappers/make-factor-en-pos.mxpost.perl is a wrapper script
to create factors for a factored translation model. You have to adapt the definition of MXPOST to point to your installation directory. TreeTagger (English, French, Spanish, German, Italian, Dutch, Bulgarian, Greek) TreeTagger14 is a tool for annotating text with part-of-speech and lemma information. Installation (Linux, check web site for other platforms): mkdir /my/installation/dir cd /my/installation/dir wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/tree-tagger-linux-3.2.tar.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/tagger-scripts.tar.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/install-tagger.sh wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/english-par-linux-3.1.bin.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/french-par-linux-3.2-utf8.bin.gz 14 http://www.ims.uni-stuttgart.de/projekte/corplex/TreeTagger/ 138 4. User Guide wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/spanish-par-linux-3.1.bin.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/german-par-linux-3.2.bin.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/italian-par-linux-3.2-utf8.bin.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/dutch-par-linux-3.1.bin.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/bulgarian-par-linux-3.1.bin.gz wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/greek-par-linux-3.2.bin.gz sh install-tagger.sh The wrapper script scripts/training/wrapper/make-pos.tree-tagger.perl creates part-ofspeech factors using TreeTagger in the format expected by Moses. The command has the required parameters -tree-tagger DIR to specify the location of your installation and -l LANGUAGE to specify the two-letter code for the language (de, fr, ...). Optional parameters are -basic to output only basic part-of-speech tags (VER instead of VER:simp -- not available for all languages), and --stem to output stems instead of part-of-speech tags. Treetagger can also shallow parse the sentence, labelling it with chunk tags. See their website15 for details. FreeLing FreeLing16 is a set of a tokenizers, morpological analyzers, syntactic parsers. and other language tools for Asturian, Catalan, English, Galician, Italian, Portuguese, Russian, Spanish, and Welsh. 4.2.4 Syntactic Parsers Collins (English) Michael Collins17 developed the first statistical parser as part of his PhD thesis. It is implemented in C. Installation: mkdir /your/installation/dir cd /your/installation/dir wget http://people.csail.mit.edu/mcollins/PARSER.tar.gz tar xzf PARSER.tar.gz cd COLLINS-PARSER/code make Collins parser also requires the installation of MXPOST (Section 4.2.3). A wrapper file to generate parse trees in the format required to train syntax models with Moses is provided in scrips/training/wrapper/parse-en-collins.perl. 15 http://www.ims.uni-stuttgart.de/projekte/corplex/TreeTagger/DecisionTreeTagger.html http://nlp.lsi.upc.edu/freeling/ 17 http://people.csail.mit.edu/mcollins/ 16 4.2. External Tools 139 BitPar (German, English) Helmut Schmid developed BitPar18 , a parser for highly ambiguous probabilistic context-free grammars (such as treebank grammars). BitPar uses bit-vector operations to speed up the basic parsing operations by parallelization. It is implemented in C and distributed as compiled code. Installation: mkdir /your/installation/dir cd /your/installation/dir wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/BitPar/BitPar.tar.gz tar xzf BitPar.tar.gz cd BitPar/src make cd ../.. You will also need the parsing model for German which was trained on the Tiger treebank: wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/BitPar/GermanParser.tar.gz tar xzf GermanParser.tar.gz cd GermanParser/src make cd ../.. There is also an English parsing model. LoPar (German) LoPar19 is an implementation of a parser for head-lexicalized probabilistic context-free grammars, which can be also used for morphological analysis. The program is distributed without source code. Installation: mkdir /my/installation/dir cd /my/installation/dir wget ftp://ftp.ims.uni-stuttgart.de/pub/corpora/LoPar/lopar-3.0.linux.tar.gz tar xzf lopar-3.0.linux.tar.gz cd LoPar-3.0 18 19 http://www.ims.uni-stuttgart.de/tcl/SOFTWARE/BitPar.html http://www.ims.uni-stuttgart.de/tcl/SOFTWARE/LoPar.html 140 4. User Guide Berkeley Parser The Berkeley is a phrase structure grammar parser implemented in Java and distributed open source. Models are provided for English, Bugarian, Arabic, Chinese, French, German. http://code.google.com/p/berkeleyparser/ 4.2.5 Other Open Source Machine Translation Systems Joshua Joshua20 is a machine translation decoder for hierarchical models. Joshua development is centered at the Center for Language and Speech Processing at the Johns Hopkins University in Baltimore, Maryland. It is implemented in Java. cdec Cdec21 is a decoder, aligner, and learning framework for statistical machine translation and other structured prediction models written by Chris Dyer in the University of Maryland Department of Linguistics. It is written in C++. Apertium Apertium22 is an open source rule-based machine translation (RBMT) system, maintained principally by the University of Alicante and Prompsit Engineering. Docent Docent23 is a decoder for phrase-based SMT that treats complete documents, rather than single sentences, as translation units and permits the inclusion of features with cross-sentence dependencies. It is developed by Christian Hardmeier and implemented in C++ 4.2.6 Other Translation Tools COSTA MT Evaluation Tool COSTA MT Evaluation Tool24 is an open-source Java program that can be used to evaluate manually the quality of the MT output. It is simple in use, designed to allow MT potential users and developers to analyse their engines using a friendly environment. It enables the ranking of the quality of MT output segment-by-segment for a particular language pair. 20 http://joshua.sourceforge.net/Joshua/Welcome.html http://cdec-decoder.org/ 22 http://www.apertium.org/ 23 https://github.com/chardmeier/docent/wiki 24 https://code.google.com/p/costa-mt-evaluation-tool/ 21 4.3. Advanced Features of the Decoder 141 Appraise Appraise25 is an open-source tool for manual evaluation of Machine Translation output. Appraise allows to collect human judgments on translation output, implementing annotation tasks such as translation quality checking, ranking of translations, error classification, and manual post-editing. It is used in the ACL WMT evaluation campaign26 . Subsection last modified on January 13, 2014, at 12:07 PM 4.3 Advanced Features of the Decoder The basic features of the decoder are explained in the Tutorial (Section 3.1). Here, we describe some additional features that have been demonstrated to be beneficial in some cases. 4.3.1 Lexicalized Reordering Models The default standard model that for phrase-based statistical machine translation is only conditioned on movement distance and nothing else. However, some phrases are reordered more frequently than others. A French adjective like extérieur typically gets switched with the preceding noun, when translated into English. Hence, we want to consider a lexicalized reordering model that conditions reordering on the actual phrases. One concern, of course, is the problem of sparse data. A particular phrase pair may occur only a few times in the training data, making it hard to estimate reliable probability distributions from these statistics. Therefore, in the lexicalized reordering model we present here, we only consider three reordering types: (m) monotone order, (s) switch with previous phrase, or (d) discontinuous. See below for an illustration of these three different types of orientation of a phrase. 25 26 https://github.com/cfedermann/Appraise http://www.statmt.org/wmt13/ 142 4. User Guide To put it more formally, we want to introduce a reordering model po that predicts an orientation type {m,s,d} given the phrase pair currently used in translation: orientation {m, s, d} po (orientation|f,e) How can we learn such a probability distribution from the data? Again, we go back to the word alignment that was the basis for our phrase table. When we extract each phrase pair, we can also extract its orientation type in that specific occurrence. Looking at the word alignment matrix, we note for each extracted phrase pair its corresponding orientation type. The orientation type can be detected, if we check for a word alignment point to the top left or to the top right of the extracted phrase pair. An alignment point to the top left signifies that the preceding English word is aligned to the preceding Foreign word. An alignment point to the top right indicates that the preceding English word is aligned to the following french word. See below for an illustration. 4.3. Advanced Features of the Decoder 143 The orientation type is defined as follows: • monotone: if a word alignment point to the top left exists, we have evidence for monotone orientation. • swap: if a word alignment point to the top right exists, we have evidence for a swap with the previous phrase. • discontinuous: if neither a word alignment point to top left nor to the top right exists, we have neither monotone order nor a swap, and hence evidence for discontinuous orientation. We count how often each extracted phrase pair is found with each of the three orientation types. The probability distribution po is then estimated based on these counts using the maximum likelihood principle: po (orientation|f,e) = count(orientation,e,f) / Σo count(o,e,f) Given the sparse statistics of the orientation types, we may want to smooth the counts with the unconditioned maximum-likelihood probability distribution with some factor σ: po (orientation) = Σf Σe count(orientation,e,f) / Σo Σf Σe count(o,e,f) po (orientation|f, e) = (σp(orientation) + count(orientation, e, f ))/(σ + Σo count(o, e, f )) There are a number of variations of this lexicalized reordering model based on orientation types: • bidirectional: Certain phrases may not only flag, if they themselves are moved out of order, but also if subsequent phrases are reordered. A lexicalized reordering model for this decision could be learned in addition, using the same method. • f and e: Out of sparse data concerns, we may want to condition the probability distribution only on the foreign phrase (f) or the English phrase (e). • monotonicity: To further reduce the complexity of the model, we might merge the orientation types swap and discontinuous, leaving a binary decision about the phrase order. 144 4. User Guide These variations have shown to be occasionally beneficial for certain training corpus sizes and language pairs. Moses allows the arbitrary combination of these decisions to define the reordering model type (e.g. bidirectional-monotonicity-f). See more on training these models in the training section of this manual. Enhanced orientation detection As explained above, statistics about the orientation of each phrase can be collected by looking at the word alignment matrix, in particular by checking the presence of a word at the top left and right corners. This simple approach is capable of detecting a swap with a previous phrase that contains a word exactly aligned on the top right corner, see case (a) in the figure below. However, this approach cannot detect a swap with a phrase that does not contain a word with such an alignment, like the case (b). A variation to the way phrase orientation statistics are collected is the so-called phrase-based orientation model by Tillmann (2004)27 , which uses phrases both at training and decoding time. With the phrase-based orientation model, the case (b) is properly detected and counted during training as a swap. A further improvement of this method is the hierarchical orientation model by Galley and Manning (2008)28 , which is able to detect swaps or monotone arrangements between blocks even larger than the length limit imposed to phrases during training, and larger than the phrases actually used during decoding. For instance, it can detect at decoding time the swap of blocks in the case (c) shown below. (Figure from Galley and Manning, 2008) Empirically, the enhanced orientation methods should be used with language pairs involving significant word re-ordering. 4.3.2 Operation Sequence Model (OSM) The Operation Sequence Model as described in Durrani et al. (2011)29 and Durrani et al. (2013)30 has been integrated into Moses. What is OSM? OSM is an N-gram-based translation and reordering model that represents aligned bilingual corpus as a sequence of operations and learns a Markov model over the resultant sequences. 27 http://www.aclweb.org/anthology-new/N/N04/N04-4026.pdf http://www.aclweb.org/anthology/D/D08/D08-1089.pdf 29 http://aclweb.org/anthology//P/P11/P11-1105.pdf 30 http://aclweb.org/anthology/P/P13/P13-2071.pdf 28 4.3. Advanced Features of the Decoder 145 Possible operations are (i) generation of a sequence of source and target words (ii) insertion of gaps as explicit target positions for reordering operations, and (iii) forward and backward jump operations which do the actual reordering. The probability of a sequence of operations is defined according to an N-gram model, i.e., the probability of an operation depends on the n-1 preceding operations. Let O = o1 , ... , oN be a sequence of operations as hypothesized by the translator to generate a word-aligned bilingual sentence pair < F;E;A >; the model is then defined as: Q posm (F,E,A) = p(o1 ,...,oN ) = i p(oi |oi-n+1 ...oi-1 ) The OSM model addresses several drawbacks of the phrase-based translation and lexicalized reordering models: i) it considers source and target contextual information across phrasal boundries and does not make independence assumption, ii) it is based on minimal translation units therefore does not have the problem of spurious phrasal segmentation, iii) it consider much richer conditioning than the lexcialized reordering model which only learns orientation of a phrase w.r.t previous phrase (or block of phrases) ignoring how previous words were translated and reordered. The OSM model conditions translation and reordering decisions on ’n’ previous translation and reordering decisions which can span across phrasal boundaries. A list of operations is given below: Generate (X,Y): X and Y are source and target cepts in an MTU (minimal translation unit). This operation causes the words in Y and the first word in X to be added to the target and source strings respectively, that were generated so far. Subsequent words in X are added to a queue to be generated later. Continue Source Cept: The source words added to the queue by the Generate (X,Y) operation are generated by the Continue Source Cept operation. Each Continue Source Cept operation removes one German word from the queue and copies it to the source string. Generate Source Only (X): The words in X are added at the current position in the source string. This operation is used to generate an target word with no corresponding target word. Generate Target Only (Y): The words in Y are added at the current position in the target string. This operation is used to generate an target word with no corresponding source word. Generate Identical: The same word is added at the current position in both the source and target strings. The Generate Identical operation is used during decoding for the translation of unknown words. Insert Gap: This operation inserts a gap which acts as a placeholder for the skipped words. There can be more than one open gap at a time. Jump Back (W): This operation lets the translator jump back to an open gap. It takes a parameter W specifying which gap to jump to. W=1 for the gap closest to the right most source word covered, W=2 for the second most closest and so on. Jump Forward: This operation makes the translator jump to the right-most source word so far covered. It is performed when the next source word to be generated is at the right of the source word generated and does not follow immediately 146 4. User Guide The example shown in figure is deterministically converted to the following operation sequence: Generate Identical -- Generate (hat investiert, invested) -- Insert Gap -- Continue Source Cept -- Jump Back (1) -- Generate (Millionen, million) -- Generate Source Only (von) -- Generate (Dollars, dollars) -Generate (in, in) -- Generate (die, the) -- Generate (Untersuchungen, research) Usage To enable the OSM model in phrase-based decoder, just put the following in the EMS config file: operation-sequence-model = "yes" operation-sequence-model-order = 5 operation-sequence-model-settings = "" Due to data sparsity the lexically driven OSM model may often fall back to very small context sizes. This problem is addressed in Durrani et al. (2014b)31 by learning operation sequences over generalized representations such as POS/Morph tags/word classes (See Section: Classbased Models). If the data has been augmented with additional factors, then use operation-sequence-model-settings = "--factor 0-0+1-1" "0-0" will learn OSM model over lexical forms and "1-1" will learn OSM model over second factor (POS/Morph/Cluster-id etc.). Note that using operation-sequence-model-settings = "" for a factor augmented training data is an error. Use operation-sequence-model-settings = "--factor 0-0" if you only intend to train OSM model over surface form in such a scenario. In case you are not using EMS and want to train OSM model manually, you will need to do two things: 1) Run the following command 31 http://www.statmt.org/OSMOSES/factor-v2.pdf 4.3. Advanced Features of the Decoder 147 /path-to-moses/scripts/OSM/OSM-Train.perl --corpus-f corpus.fr --corpus-e corpus.en --alignment aligned.grow-diag-final-and --order 5 --out-dir /path-to-experiment/model/OSM 2) Edit model/moses.ini to add [feature] ... OpSequenceModel name=OpSequenceModel0 num-features=5 path=/path-to-experiment/model/OSM/operat ... [weight] ... OpSequenceModel0= 0.08 -0.02 0.02 -0.001 0.03 ... 4.3.3 Class-based Models Automatically clustering the training data into word classes in order to obtain smoother distributions and better generalizations has been a widely known and applied technique in natural language processing. Using class-based models have shown to be useful when translating into morphologically rich languages. We use the mkcls utility in GIZA to cluster source and target vocabularies into classes. This is generally run during alignment process where data is divided into 50 classes to estimate IBM Model-4. Durrani et al. (2014b)32 found using different number of clusters to be useful for different language pairs. To map the data (say corpus.fr) into higher number of clusters (say 1000) use: /path-to-GIZA/statmt/bin/mkcls Űc1000 -n2 -p/path-to-corpus/corpus.fr -V/path-to-experiment/training/prepared.stepID/fr.vcb.classes opt To annotate the data with cluster-ids add the following to the EMS-config file: ################################################################# # FACTOR DEFINITION [INPUT-FACTOR] temp-dir =working-dir/training/factor
[INPUT-FACTOR:mkcls]
### script that generates this factor
#
factor-script = "/path-to-moses/scripts/training/wrappers/make-factor-brown-cluster-mkcls.perl 0 $working-dir/training/prepared.stepID/$input-extension.vcb.classes"
[OUTPUT-FACTOR:mkcls]
### script that generates this factor
#
factor-script = "/path-to-moses/scripts/training/wrappers/make-factor-brown-cluster-mkcls.perl 0 $working-dir/training/prepared.stepID/$output-extension.vcb.classes"
#################################################################
32
http://www.statmt.org/OSMOSES/factor-v2.pdf
148
4. User Guide
Adding the above will augment the training data with cluster-ids. These can be enabled in
different models. For example to train a joint-source target phrase-translation model, add the
following to the EMS-config file:
[TRAINING]
input-factors = word mkcls
output-factors = word mkcls
alignment-factors = "word -> word"
translation-factors = "word+mkcls -> word+mkcls"
reordering-factors = "word -> word"
decoding-steps = "t0"
To train a target sequence model over cluster-ids, add the following to the EMS config-file
[LM]
[LM:corpus-mkcls]
raw-corpus = /path-to-raw-monolingual-data/rawData.en
factors = mkcls
settings = "-unk"
To train operation sequence model over cluster-ids, use the following in the EMS config-file
[TRAINING]
operation-sequence-model-settings = "--factor 1-1"
if you want to train both lexically driven and class-based OSM models then use:
[TRAINING]
operation-sequence-model-settings = "--factor 0-0+1-1"
4.3.
4.3.4
149
Unsupervised Transliteration Model
Character-based translation model/Transliteration models have shown to be quite useful in
MT for translating OOV words, for disambiguation and for translating closely related languages. A transliteration module as described in Durrani et al. (2014a)33 has been integrated
into Moses. It is completely unsupervised and language independent. It extracts transliteration
corpus from the parallel data and builds a transliteration model from it which can then be used
to translate OOV word or named-entities.
Specification with experiment.perl
To enable transliteration module add the following to the EMS config file:
[TRAINING]
transliteration-module = "yes"
It will extract transliteration corpus from the word-aligned parallel data and learn a characterbased model from it.
To use the post-decoding transliteration (Method 2 as described in the paper) add the following
lines
post-decoding-transliteration = "yes"
To use the in-decoding method (Method 3 as described in the paper) add the following lines
in-decoding-transliteration = "yes"
transliteration-file = /file containing list of words to be transliterated/
Post-decoding method obtains the list of OOV words automatically by running the decoder.
The in-decoding method requires the user to provide the list of words to be transliterated. This
gives a freedom to transliterate any additional words that might be known to the translation
model but can also be transliterated in some scenarios. For example "Little" can be translated to
<img src=’http://www.statmt.org/moses/img/choti.png’ alt=” title=” /> in Urdu when it is
used as adjective and transliterated to <img src=’http://www.statmt.org/moses/img/Little.png’
alt=” title=” /> when it is a name as in "Stuart Little". You can add the OOV list as obtained
from Method 2 if you don’t want to add any other words. Transliterating all the words in
the test-set might be helpful when translating between closely related language pairs such as
Hindi-Urdu, Thai-Lao etc. See Durrani and Koehn (2014)34 for a case-study.
33
34
http://aclweb.org/anthology//E/E14/E14-4029.pdf
http://www.statmt.org/OSMOSES/interpolateTransliterate-v2.pdf
150
4. User Guide
Steps for use outside experiment.perl
Execute command to train transliteration:
../mosesdecoder/scripts/Transliteration/train-transliteration-module.pl \
--corpus-f <foreign text> --corpus-e <target text> \
--alignment <path to aligned text> \
--moses-src-dir <moses decoder path> --external-bin-dir <external tools> \
--input-extension <input extension>--output-extension <output-extension> \
--srilm-dir <sri lm binary path> --out-dir <path to generate output files>
Train moses with transliteration option on
nohup nice train-model.perl -root-dir train -corpus <path to parallel corpus> \
-f <foreign> -e <target> -alignment grow-diag-final-and \
-reordering msd-bidirectional-fe -lm 0:3:<path to lm>:8 \
-external-bin-dir <external tools> -post-decoding-translit yes \
-transliteration-phrase-table <path to transliteration phrase table> >& training.out &
First pass decoding to generate output oov file and translation output without transliteration
nohup nice <path to moses> -f <moses.ini file> \
-output-unknowns <path to oov file to be output> \
< <path to test input> > <path to test output> 2> <path to trace output>
Second pass decoding to transliterate to the output
./post-decoding-transliteration.pl --moses-src-dir <moses decoder> \
--external-bin-dir <external tools> --transliteration-model-dir <transliteration model>
--oov-file <oov file obtained in previous step> \
--input-file <translated file obtained in previous step> \
--output-file <output translated file> \
--input-extension <foreign> --output-extension <english> \
--language-model <path to language model> \
--decoder <moses executable>
\
4.3.
4.3.5
151
For larger tasks the phrase tables usually become huge, typically too large to fit into memory.
Therefore, Moses supports a binary phrase table with on-demand loading, i.e. only the part of
the phrase table that is required to translate a sentence is loaded into memory.
There are currently 3 binary formats to do this:
• OnDisk phrase-table. Works with SCFG models and phrase-based models.
• Binary phrase-table. Works with phrase-based models only.
• Compress phrase-table. Works with phrase-based models only (may be extended in the
near future). Small and fast. Described here35 .
On-Disk Phrase table
This phrase-table can be used for both phrase-based models and hierarchical models. (It can be
used for fully syntax models too, but is likely to be very slow).
You first need to convert the rule table into a binary prefix format. This is done with the command CreateOnDiskPt:
CreateOnDiskPt [#source factors] [#target factors] [#scores] [ttable-limit]
[index of p(e|f) (usually 2)] [input text pt] [output directory]
\
e.g.
~/CreateOnDiskPt 1 1 4 100 2 phrase-table.1.gz phrase-table.1.folder
This will create a directory, phrase-table.1.folder, with the following files:
Misc.dat
Source.dat
TargetColl.dat
TargetInd.dat
Vocab.dat
The configuration file moses.ini should also be changed so that the binary files is used instead
of the text file. You should change it from:
35
152
4. User Guide
[feature]
PhraseDictionaryMemory path=phrase-table.1.gz ....
to
[feature]
PhraseDictionaryOnDisk path=phrase-table.1.folder ....
Binary Phrase table
NB- Works with phrase-based models only. NB2 - This phrase-table is now included for
bacward-compatibility. It may be deleted in future.
You have to convert the standard ASCII phrase tables into the binary format. Here is an example (standard phrase table phrase-table, with 4 scores):
cat phrase-table | LC_ALL=C sort | bin/processPhraseTable \
-ttable 0 0 - -nscores 4 -out phrase-table
Options:
• -ttable int int string -- translation table file, use ’-’ for stdin
• -out string -- output file name prefix for binary translation table
• -nscores int -- number of scores in translation table
If you just want to convert a phrase table, the two integers in the -ttable option do not matter,
so use 0’s.
Important: If your data is encoded in UTF8, make sure you set the environment variable with
the LC_ALL=C before sorting. If your phrase table is already sorted, you can skip that.
The output files will be:
phrase-table.binphr.idx
phrase-table.binphr.srctree
phrase-table.binphr.srcvoc
phrase-table.binphr.tgtdata
phrase-table.binphr.tgtvoc
4.3.
153
In the Moses configuration file, specify only the file name stem phrase-table as phrase table
and set the type to 1, i.e.:
[feature]
PhraseDictionaryBinary path=phrase-table ...
Word-to-word alignment:
There are 2 arguments to the decoder that enables it to print out the word alignment information
-alignment-output-file [file]
print out the word alignment for the best translation to a file.
-print-alignment-info-in-n-best
print the word alignment information of each entry in the n-best list as an extra column in the
n-best file.
Word alignment is included in the phrase-table by default (as of November 2012). To exclude
--NoWordAlignment
as an argument to the score program.
When binarizing the phrase-table, the word alignment is also included by default. To turn this
behaviour off for the phrase-based binarizer:
processPhraseTable -no-alignment-info ....
Or
154
4. User Guide
processPhraseTableMin -no-alignment-info ....
(For the compact phrase-table representation).
There is no way to exclude word alignment information from the chart-based binarization process.
Phrase-based binary format When word alignment information is stored, the two output files
".srctree" and " .tgtdata" will end with the suffix ".wa".
Note: The argument
-use-alignment-info
-print-alignment-info
has been deleted from the decoder. -print-alignment-info did nothing. -use-alignment-info
is now inferred from the arguments
-alignment-output-file
-print-alignment-info-in-n-best
-include-alignment-in-n-best
has been renamed
-include-segmentation-in-n-best
to reflect what it actually does.
The word alignment MUST be enabled during binarization, otherwise the decoder will
1. complain
2. carry on blindly but doesn’t print any word alignment
4.3.
4.3.6
155
The reordering tables may be also converted into a binary format. The command is slightly
simpler:
mosesdecoder/bin/processLexicalTable -in reordering-table -out reordering-table
The file names for input and output are typically the same, since the actual output file names
have similar extensions to the phrase table file names.
4.3.7
Compact Phrase Table
A compact phrase table implementation is available that is around 6 to 7 times smaller and
than the original binary phrase table. It can be used in-memory and for on-demand loading.
Like the original phrase table, this can only be used for phrase-based models. If you use this or
the compact lexical reordering table below, please cite:
• Marcin Junczys-Dowmunt: Phrasal Rank-Encoding: Exploiting Phrase Redundancy
and Translational Relations for Phrase Table Compression36 , Proceedings of the Machine Translation Marathon 2012, The Prague Bulletin of Mathematical Linguistics, vol.
98, pp. 63-74, 2012.
to the included instructions. Make sure the installation target directory contains an "include"
and a "lib" directory. Next you need to recompile Moses with
./bjam --with-cmph=/path/to/cmph
Now, you can convert the standard ASCII phrase tables into the compact format. Phrase tables
are required to be sorted as above. For a maximal compression effect, it is advised to generate a phrase table with phrase-internal word alignment (this is the default). If you want to
compress a phrase table without alignment information, rather use -encoding None (see advanced options below). It is possible to use the default encoding (PREnc) without alignment
information, but it will take much longer. For now, there may be problems to compact phrase
tables on 32-bit systems since virtual memory usage quickly exceeds the 3 GB barrier.
Here is an example (standard phrase table phrase-table, with 5 scores) which produces a single
file phrase-table.minphr:
mosesdecoder/bin/processPhraseTableMin -in phrase-table.gz -out phrase-table -nscores 4 -threads 4
36
http://ufal.mff.cuni.cz/pbml/98/art-junczys-dowmunt.pdf
156
4. User Guide
In the Moses config file, specify the WHOLE file name of the phrase table:
[feature]
PhraseDictionaryCompact path=phrase-table.minphr ...
Options:
• -in string -- input table file name
• -out string -- prefix of binary table file
• -nscores int -- number of score components in phrase table
• -no-alignment-info -- do not include alignment info in the binary phrase table
• -T string -- path to custom temporary directory
As for the original phrase table, the option -no-alignment-info omits phrase internal alignment information in the phrase table and should also be used if you provide a phrase table
without alignment information in the phrase table. Also if no alignment data is given in
the phrase table you should use -encoding None (see below), since the default compression
method assumes that alignment information is present.
Since compression is quite processor-heavy, it is advised to use the -threads option to increase
speed.
Advanced options: Default settings should be fine for most of your needs, but the size of the
phrase table can be tuned to your specific needs with the help of the advanced options.
Options:
• -encoding string -- encoding type: PREnc REnc None (default PREnc)
• -rankscore int -- score index of P(t|s) (default 2)
• -maxrank int -- maximum rank for PREnc (default 100)
• -landmark int -- use landmark phrase every 2n̂ source phrases (default 10)
• -fingerprint int -- number of bits used for source phrase fingerprints (default 16)
• -join-scores -- single set of Huffman codes for score components
• -quantize int -- maximum number of scores per score component
• -no-warnings -- suppress warnings about missing alignment data
Encoding methods: There are two encoding types that can be used on-top the standard compression methods, Phrasal Rank-Encoding (PREnc) and word-based Rank Encoding (REnc).
PREnc (see Junczys-Dowmunt (MT Marathon 2012)37 for details) is used by default and requires a phrase table with phrase-internal alignment to reach its full potential. PREnc can also
work without explicit alignment information, but encoding is slower and the resulting file will
be bigger, but smaller than without PREnc. The tool will warn you about every line that misses
alignment information if you use PREnc or REnc. These warnings can be suppressed with
-no-warnings. If you use PREnc with non-standard scores you should specify which score
type is used for sorting with -rankscore int. By default this is P(t|s) which in the standard
phrase table is the third score (index 2).
37
http://ufal.mff.cuni.cz/pbml/98/art-junczys-dowmunt.pdf
4.3.
157
Basically with PREnc around, there is no reason to use REnc unless you really want to. It
requires the lexical translation table lex.f2e to be present in the same directory as the text
version phrase table. If no alignment information is available it falls back to None (See JunczysDowmunt (EAMT 2012)38 for details on REnc and None).
None is the fasted encoding method, but produces the biggest files. Concerning translation
speed, there is virtually no difference between the encoding methods when the phrase tables
are later used with Moses, but smaller files result in lesser memory-usage, especially if the
phrase tables are loaded entirely in-memory.
Indexing options: The properties of the source phrase index can be modified with the -landmark
and -fingerprint options, changing these options can affect file size and translation quality,
so do it carefully. Junczys-Dowmunt (TSD 2012)39 contains a discussion of these values and
their effects.
Scores and quantization: You can reduce the file size even more by using score quantization.
E.g. with -quantize 1000000, a phrase table is generated that uses at most one million different scores for each score type. Be careful, low values will affect translation quality. By default,
each score type is encoded with an own set of Huffman codes, with the -join-scores option
only one set is used. If this option is combined with -quantize N, the summed number of
different scores for all scores types will not exceed N.
phrase table directly into memory at start up. Without this option, on-demand loading is used
by default.
4.3.8
Compact Lexical Reordering Table
The compact lexical reordering table produces files about 12 to 15 times smaller than the original Moses binary implementation. As for the compact phrase table you need to install CMPH
and link against it. Reordering tables must be sorted in the same way as the phrase tables
above. The command below produces a single file reordering-table.minlexr.
mosesdecoder/bin/processLexicalTableMin -in reordering-table.gz -out reordering-table -threads 4
If you include the prefix in the Moses config file, the compact reordering table will be recognized and loaded automatically. You can start Moses with the option -minlexr-memory to load
the compact lexical reordering table directly into memory at start up.
Options: See the compact phrase table above for a description of available common options.
38
39
http://hltshare.fbk.eu/EAMT2012/html/Papers/57.pdf
http://www.staff.amu.edu.pl/~junczys/dokuwiki/lib/exe/fetch.php?cache=&media=wiki:mjd2012tsd1.pdf
158
4. User Guide
4.3.9
XML Markup
Sometimes we have external knowledge that we want to bring to the decoder. For instance, we
might have a better translation system for translating numbers of dates. We would like to plug
in these translations to the decoder without changing the model.
The -xml-input flag is used to activate this feature. It can have one of four values:
• exclusive Only the XML-specified translation is used for the input phrase. Any phrases
from the phrase table that overlap with that span are ignored.
• inclusive The XML-specified translation competes with all the phrase table choices for
that span.
• constraint The XML-specified translation competes with phrase table choices that contain the specified translation.
• ignore The XML-specified translation is ignored completely.
• pass-through (default) For backwards compatibility, the XML data is fed straight through
to the decoder. This will produce erroneous results if the decoder is fed data that contains
XML markup.
The decoder has an XML markup scheme that allows the specification of translations for parts
of the sentence. In its simplest form, we can tell the decoder what to use to translate certain
words or phrases in the sentence:
% echo ’das ist <np translation="a cute place">ein kleines haus</np>’ \
| moses -xml-input exclusive -f moses.ini
this is a cute place
% echo ’das ist ein kleines <n translation="dwelling">haus</n>’ \
| moses -xml-input exclusive -f moses.ini
this is a little dwelling
The words have to be surrounded by tags, such as <np...> and </np>. The name of the tags
can be chosen freely. The target output is specified in the opening tag as a parameter value for
a parameter that is called english for historical reasons (the canonical target language).
We can also provide a probability along with these translation choice. The parameter must be
named prob and should contain a single float value. If not present, an XML translation option
is given a probability of 1.
% echo ’das ist ein kleines <n translation="dwelling" prob="0.8">haus</n>’ \
| moses -xml-input exclusive -f moses.ini \
this is a little dwelling
This probability isn’t very useful without letting the decoder have other phrase table entries
"compete" with the XML entry, so we switch to inclusive mode. This allows the decoder to
use either translations from the model or the specified xml translation:
4.3.
159
% echo ’das ist ein kleines <n translation="dwelling" prob="0.8">haus</n>’ \
| moses -xml-input inclusive -f moses.ini
this is a small house
The switch -xml-input inclusive gives the decoder a choice between using the specified
translations or its own. This choice, again, is ultimately made by the language model, which
takes the sentence context into account.
This doesn’t change the output from the non-XML sentence because that prob value is first
logged, then split evenly among the number of scores present in the phrase table. Additionally,
in the toy model used here, we are dealing with a very dumb language model and phrase table.
Setting the probability value to something astronomical forces our option to be chosen.
% echo ’das ist ein kleines <n translation="dwelling" prob="0.8">haus</n>’ \
| moses -xml-input inclusive -f moses.ini
this is a little dwelling
Multiple translation can be specified if separated by two bars (||):
% echo ’das ist ein kleines <n translation="dwelling||house" prob="0.8||0.2">haus</n>’ \
| moses -xml-input inclusive -f moses.ini
The XML-input implementation is NOT currently compatible with factored models or confusion networks.
Options:
• -xml-input (’pass-through’ (default), ’inclusive’, ’constraint’, ’exclusive’,
’ignore’)
4.3.10
Generating n-Best Lists
The generation of n-best lists (the top n translations found by the search according to the model)
is pretty straight-forward. You simple have to specify the file where the n-best list will be stored
and the size of the n-best list for each sentence.
Example: The command
% moses -f moses.ini -n-best-list listfile 100 < in
160
4. User Guide
stores the n-best list in the file listfile with up to 100 translations per input sentence.
Here an example n-best list:
0 ||| we must discuss on greater vision . ||| d: 0 -5.56438 0 0 -7.07376 0 0 \
lm: -36.0974 -13.3428 tm: -39.6927 -47.8438 -15.4766 -20.5003 4.99948 w: -7 ||| -9.2298
0 ||| we must also discuss on a vision . ||| d: -10 -2.3455 0 -1.92155 -3.21888 0 -1.51918 \
lm: -31.5841 -9.96547 tm: -42.3438 -48.4311 -18.913 -20.0086 5.99938 w: -8 ||| -9.26197
0 ||| it is also discuss a vision . ||| d: -10 -1.63574 -1.60944 -2.70802 -1.60944 -1.94589 -1.08417 \
lm: -31.9699 -12.155 tm: -40.4555 -46.8605 -14.3549 -13.2247 4.99948 w: -7 ||| -9.31777
Each line of the n-best list file is made up of (separated by |||):
•
•
•
•
sentence number (in above example 0, the first sentence)
output sentence
individual component scores (unweighted)
weighted overall score
Note that it is possible (and very likely) that the n-best list contains many sentences that look
the same on the surface, but have different scores. The most common reason for this is different phrase segmentation (two words may be mapped by a single phrase mapping, or by two
individual phrase mappings for each word).
To produce an n-best list that only contains the first occurrence of an output sentence, add the
word distinct after the file and size specification:
% moses -f moses.ini -n-best-list listfile 100 distinct < in
This creates an n-best list file that contains up to 100 distinct output sentences for each input
sentences. Note that potentially a large numbers of candidate translations have to be examined
to find the top 100. To keep memory usage in check only 20 times the specified number of
distinct entries are examined. This factor can be changed with the switch -n-best-factor.
Options:
• -n-best-list FILE SIZE [distinct] -– output an n-best list of size SIZE to file FILE
• -n-best-factor FACTOR -– exploring at most FACTOR*SIZE candidates for distinct
• -include-alignment-in-n-best -– output of word-to-word alignments in the nbest list; it requires that w2w alignments are included in the phrase table and that
-use-alignment-info is set. (See here (Section 4.3.11) for further details).
4.3.11
Word-to-word alignment
If the phrase table (binary or textual) includes word-to-word alignments between source and
target phrases (see "Score Phrases" (Section 5.9) and "Binary Phrase Table" (Section 4.3.5)),
Moses can report them in the output.
4.3.
161
There are four options that control the output of alignment information: -use-alignment-info,
-print-alignment-info, -print-alignment-info-in-n-best, and -alignment-output-file.
For instance, by translating the sentence "ich frage" from German into English and activating
all parameters, you get in the verbose output:
[total=-1.429] <<features>> [f2e: 0=0 1=1] [e2f: 0=0 1=1]
The last two fields report the word-to-word alignments from source to target and from target
to source, respectively.
In the n-best list you get:
0
0
0
0
|||
|||
|||
|||
i
i
i
i
ask ||| ...feature_scores.... ||| -1.42906 ||| 0-1=0-1 ||| 0=0 1=1 ||| 0=0 1=1
am asking ||| ...feature_scores.... ||| -2.61281 ||| 0-1=0-2 ||| 0=0 1=1,2 ||| 0=0 1=1 2=1
ask you ||| ...feature_scores.... ||| -3.1068 ||| 0-1=0-2 ||| 0=0 1=1,2 ||| 0=0 1=1 2=1
ask this ||| ...feature_scores.... ||| -3.48919 ||| 0-1=0-2 ||| 0=0 1=1 ||| 0=0 1=1 2=-1
Indexes (starting from 0) are used to refer to words. ’2=-1’ means that the word of index 2 (i.e.
the word) is not associated with any word in the other language. For instance, by considering
the last translation hypothesis "i ask this" of "ich frage", the source to target alignment ("0=0
1=1") means that:
German
ich
frage
-> English
-> i
and vice versa the target to source alignment ("0=0 1=1 2=-1") means that:
English
i
this
-> German
-> ich
-> frage
->
Note: in the same translation hypothesis, the the field "0-1=0-2" after the global score refers
to the phrase-to-phrase alignment and means that "ich frage" is translated as a phrase into the
three-word English phrase "i ask you".
This information is generated if the option -include-alignment-in-n-best is activated.
162
4. User Guide
Important: the phrase table can include different word-to-word alignments for the sourceto-target and target-to-source directions, at least in principle. Hence, the two alignments can
differ.
Options:
• -use-alignment-info -- to activate this feature (required for binarized ttables, see
"Binary Phrase Table" (Section 4.3.5)).
• -print-alignment-info -- to output the word-to-word alignments into the verbose
output.
• -print-alignment-info-in-n-best -- to output the word-to-word alignments into
the n-best lists.
• -alignment-output-file outfilename -- to output word-to-word alignments into a
separate file in a compact format (one line per sentence).
4.3.12
Minimum Bayes Risk Decoding
Minumum Bayes Risk (MBR) decoding was proposed by Kumar and Byrne (HLT/NAACL
2004)40 . Roughly speaking, instead of outputting the translation with the highest probability,
MBR decoding outputs the translation that is most similar to the most likely translations. This
requires a similarity measure to establish similar. In Moses, this is a smoothed BLEU score.
Using MBR decoding is straight-forward, just use the switch -mbr when invoking the decoder.
Example:
% moses -f moses.ini -mbr < in
MBR decoding uses by default the top 200 distinct candidate translations to find the translation
with minimum Bayes risk. If you want to change this to some other number, use the switch
-mbr-size:
% moses -f moses.ini -decoder-type 1 -mbr-size 100 < in
MBR decoding requires that the translation scores are converted into probabilities that add up
to one. The default is to take the log-scores at face value, but you may get better results with
scaling the scores. This may be done with the switch -mbr-scale, so for instance:
% moses -f moses.ini -decoder-type 1 -mbr-scale 0.5 < in
40
http://mi.eng.cam.ac.uk/~wjb31/ppubs/hlt04_mbr_smt.pdf
4.3.
163
Options:
• -mbr -- use MBR decoding
• -mbr-size SIZE -- number of translation candidates to consider (default 200)
• -mbr-scale SCALE -- scaling factor used to adjust the translation scores (default 1.0)
Note: MBR decoding and its variants is currently only implemented for the phrase-based decoder, not the chart decoder.
4.3.13
Lattice MBR and Consensus Decoding
These are extensions to MBR which may run faster or give better results. For more details see
Tromble et al (2008)41 , Kumar et al (2009)42 and De Nero et al (2009)43 . The n-gram posteriors
(required for Lattice MBR) and the ngram expectations (for Consensus decoding) are both calculated using an algorithm described in De Nero et al (2010)44 . Currently both lattice MBR and
consensus decoding are implemented as n-best list rerankers, in other words the hypothesis
space is an n-best list (not a lattice).
Here’s the list of options which affect both Lattice MBR and Consensus decoding.
Options:
• -lmbr -- use Lattice MBR decoding
• -con -- use Consensus decoding
• -mbr-size SIZE -- as for MBR
• -mbr-scale SCALE -- as for MBR
• -lmbr-pruning-factor FACTOR -- mean words per node in pruned lattice, as described in Tromble et al (2008) (default 30)
Lattice MBR has several further parameters which are described in the Tromble et al 2008 paper.
Options:
• -lmbr-p P -- The unigram precision (default 0.8)
• -lmbr-r R -- The ngram precision ratio (default 0.6)
• -lmbr-thetas THETAS Instead of specifying p and r, lattice MBR can be configured by
specifying all the ngram weights and the length penalty (5 numbers). This is described
fully in the references.
• -lmbr-map-weight WEIGHT The weight given to the map hypothesis (default 0)
Since Lattice MBR has so many parameters, a utility to perform a grid search is provided. This
is in moses-cmd/src and is called lmbrgrid. A typical usage would be
% ./lmbrgrid -lmbr-p 0.4,0.6,0.8 -lmbr-r 0.4,0.6,0.8 -mbr-scale 0.1,0.2,0.5,1 -lmbr-pruning-factor
30 -mbr-size 1000 -f moses.ini -i input.txt
41
http://www.aclweb.org/anthology-new/D/D08/D08-1065.pdf
http://www.aclweb.org/anthology-new/P/P09/P09-1019.pdf
43
http://www.aclweb.org/anthology-new/P/P09/P09-1064.pdf
44
http://www.eecs.berkeley.edu/~denero/research/papers/naacl10_denero_combination.pdf
42
\
164
4. User Guide
In other words, the same Lattice MBR parameters as for Moses are used, but this time a comma
separated list can be supplied. Each line in the output takes the following format:
<sentence-id> |||
<r> <pruning-factor> <scale> ||| <translation>
In the Moses Lattice MBR experiments that have been done to date, lattice MBR showed small
overall improvements on a NIST Arabic data set (+0.4 over map, +0.1 over MBR), once the
parameters were chosen carefully. Parameters were optimized by grid search on 500 sentences
of held-out, and the following were found to be optimal
-lmbr-p 0.8 -lmbr-r 0.8 -mbr-scale 5 -lmbr-pruning-factor 50
4.3.14
Handling Unknown Words
Unknown words are copied verbatim to the output. They are also scored by the language
model, and may be placed out of order. Alternatively, you may want to drop unknown words.
To do so add the switch -drop-unknown.
When translating between languages that use different writing sentences (say, Chinese-English),
dropping unknown words results in better BLEU scores. However, it is misleading to a human
reader, and it is unclear what the effect on human judgment is.
Options:
• -drop-unknown -- drop unknown words instead of copying them into the output
4.3.15
Output Search Graph
It may be useful for many downstream applications to have a dump of the search graph, for
instance to compile a word lattice. One the one hand you can use the -verbose 3 option,
which will give a trace of all generated hypotheses, but this creates logging of many hypotheses
that get immediately discarded. If you do not want this, a better option is using the switch
-output-search-graph FILE, which also provides some additional information.
The generated file contains lines that could be seen as both a dump of the states in the graph
and the transitions in the graph. The state graph more closely reflects the hypotheses that are
generated in the search. There are three types of hypotheses:
• The initial empty hypothesis is the only one that is not generated by a phrase translation
0 hyp=0 stack=0 [...]
4.3.
165
• Regular hypotheses
0 hyp=17 stack=1 back=0 score=-1.33208 [...] covered=0-0 out=from now on
• Recombined hypotheses
0 hyp=5994 stack=2 back=108 score=-1.57388 [...] recombined=13061 [...] covered=2-2 out=be
The relevant information for viewing each line as a state in the search graph is the sentence
number (initial 0), the hypothesis id (hyp), the stack where the hypothesis is placed (same
as number of foreign words covered, stack), the back-pointer to the previous hypotheses
(back), the score so far (score), the last output phrase (out) and that phrase’s foreign coverage
(covered). For recombined hypotheses, also the superior hypothesis id is given (recombined).
The search graph output includes additional information that is computed after the fact. While
the back-pointer and score (back, score) point to the cheapest path and cost to the beginning
of the graph, the generated output also included the pointer to the cheapest path and score
(forward, fscore) to the end of the graph.
One way to view the output of this option is a reflection of the search and all (relevant) hypotheses that are generated along the way. But often, we want to generate a word lattice, where the
states are less relevant, but the information is in the transitions from one state to the next, each
transition emitting a phrase at a certain cost. The initial empty hypothesis is irrelevant here, so
we need to consider only the other two hypothesis types:
• Regular hypotheses
0 hyp=17 [...] back=0 [...] transition=-1.33208 [...] covered=0-0 out=from now on
• Recombined hypotheses
0 [...] back=108 [...] transition=-0.640114 recombined=13061 [...] covered=2-2 out=be
For the word lattice, the relevant information is the cost of the transition (transition), its
output (out), maybe the foreign coverage (covered), and the start (back) and endpoint (hyp).
Note that the states generated by recombined hypothesis are ignored, since the transition points
to the superior hypothesis (recombined).
Here, for completeness sake, the full lines for the three examples we used above:
166
4. User Guide
0 hyp=0 stack=0 forward=9 fscore=-107.279
0 hyp=17 stack=1 back=0 score=-1.33208 transition=-1.33208 \
forward=517 fscore=-106.484 covered=0-0 out=from now on
0 hyp=5994 stack=2 back=108 score=-1.57388 transition=-0.640114 \
recombined=13061 forward=22455 fscore=-106.807 covered=2-2 out=be
When using the switch -output-search-graph-extended (or short: -osgx), a detailed score
breakdown is provided for each line. The format is the same as in the n-best list.
What is the difference between the search graph output file generated with this switch and the
true search graph?
• It contains the additional forward costs and forward paths.
• It also only contains the hypotheses that are part of a fully connected path from the initial
empty hypothesis to a final hypothesis that covers the full foreign input sentence.
• The recombined hypotheses already point to the correct superior hypothesis, while the
-verbose 3 log shows the recombinations as they happen (recall that momentarily superior hypotheses may be recombined to even better ones down the road).
Note again that you can get the full search graph with the -verbose 3 option. It is, however,
much larger and mostly consists of discarded hypotheses.
Options:
• -output-search-graph FILE -- output the search graph for each sentence in a file
• -output-search-graph-extended FILE -- output the search graph for each sentence
in a file, with detailed feature breakdown
4.3.16
During the beam search, many hypotheses are created that are too bad to be even entered on a
stack. For many of them, it is even clear before the construction of the hypothesis that it would
be not useful. Early discarding of such hypotheses hazards a guess about their viability. This is
based on correct score except for the actual language model costs which are very expensive to
compute. Hypotheses that, according to this estimate, are worse than the worst hypothesis of
the target stack, even given an additional specified threshold as cushion, are not constructed at
all. This often speeds up decoding significantly. Try threshold factors between 0.5 and 1.
Options:
the specified threshold (default: 0 = not used)
4.3.17
Maintaining stack diversity
The beam search organizes and compares hypotheses based on the number of foreign words
they have translated. Since they may have different foreign words translated, we use future
score estimates about the remaining sentence translation score.
4.3.
167
Instead of comparing such apples and oranges, we could also organize hypotheses by their
exact foreign word coverage. The disadvantage of this is that it would require an exponential
number of stacks, but with reordering limits the number of stacks is only exponential with
regard to maximum reordering distance.
Such coverage stacks are implemented in the search, and their maximum size is specified with
the switch -stack-diversity (or -sd), which sets the maximum number of hypotheses per
coverage stack.
The actual implementation is a hybrid of coverage stacks and foreign word count stacks: the
stack diversity is a constraint on which hypotheses are kept on the traditional stack. If the stack
diversity limits leave room for additional hypotheses according to the stack size limit (specified
by -s, default 200), then the stack is filled up with the best hypotheses, using score so far and
the
future score estimate.
Options:
• -stack-diversity LIMIT -- keep a specified number of hypotheses for each foreign
word coverage (default: 0 = not used)
4.3.18
Cube Pruning
Cube pruning, as described by Huang and Chiang (2007)45 , has been implemented in the Moses
decoder. This is in addition to the traditional search algorithm. The code offers developers the
opportunity to implement different search algorithms using an extensible framework.
Cube pruning is faster than the traditional search at comparable levels of search errors. To get
faster performance than the default Moses setting at roughly the same performance, use the
parameter settings:
-search-algorithm 1 -cube-pruning-pop-limit 2000 -s 2000
This uses cube pruning (-search-algorithm) that adds 2000 hypotheses to each stack (-cube-pruning-pop-limit
2000) and also increases the stack size to 2000 (-s 2000). Note that with cube pruning, the size
of the stack has little impact on performance, so it should be set rather high. The speed/quality
trade-off is mostly regulated by the cube pruning pop limit, i.e. the number of hypotheses
Stacks are organized by the number of foreign words covered, so they may differ by which
words are covered. You may also require that a minimum number of hypotheses is added for
each word coverage (they may be still pruned out, however). This is done using the switch
-cube-pruning-diversity MINIMUM which sets the minimum. The default is 0.
Options:
• -search-algorithm 1 -- turns on cube pruning
• -cube-pruning-pop-limit LIMIT -- number of hypotheses added to each stack
• -cube-pruning-diversity MINIMUM -- minimum number of hypotheses from each
coverage pattern
45
http://www.aclweb.org/anthology/P/P07/P07-1019.pdf
168
4.3.19
4. User Guide
Specifying Reordering Constraints
For various reasons, it may be useful to specify reordering constraints to the decoder, for instance because of punctuation. Consider the sentence:
I said " This is a good idea . " , and pursued the plan .
The quoted material should be translated as a block, meaning that once we start translating
some of the quoted words, we need to finish all of them. We call such a block a zone and allow
the specification of such constraints using XML markup.
I said <zone> " This is a good idea . " </zone> , and pursued the plan .
Another type of constraints are walls which are hard reordering constraints: First all words
before a wall have to be translated, before words afterwards are translated. For instance:
This is the first part . <wall /> This is the second part .
Walls may be specified within zones, where they act as local walls, i.e. they are only valid
within the zone.
I said <zone> " <wall /> This is a good idea . <wall /> " </zone> , and pursued the plan .
If you add such markup to the input, you need to use the option -xml-input with either
exclusive or inclusive (there is no difference between these options in this context).
Specifying reordering constraints around punctuation is often a good idea.
The switch -monotone-at-punctuation introduces walls around the punctuation tokens ,.!?:;".
Options:
• walls and zones have to specified in the input using the tags <zone>, </zone>, and
<wall>.
• -xml-input -- needs to be exclusive or inclusive
• -monotone-at-punctuation (-mp) -- adds walls around punctuation ,.!?:;".
4.3.
4.3.20
169
Multiple Translation Tables and Back-off Models
Moses allows the use of multiple translation tables, but there are two different ways how they
are used:
• both translation tables are used for scoring: This means that every translation option is
collected from each table and scored by each table. This implies that each translation
option has to be contained in each table: if it is missing in one of the tables, it can not be
used.
• either translation table is used for scoring: Translation options are collected from one
table, and additional options are collected from the other tables. If the same translation
option (in terms of identical input phrase and output phrase) is found in multiple tables,
separate translation options are created for each occurrence, but with different scores.
In any case, each translation table has its own set of weights.
First, you need to specify the translation tables in the section [feature] of the moses.ini configuration file, for instance:
[feature]
PhraseDictionaryMemory path=/my-dir/table1 ...
PhraseDictionaryMemory path=/my-dir/table2 ...
Secondly, you need to set weights for each phrase-table in the section [weight].
Thirdly, you need to specify how the tables are used in the section [mapping]. As mentioned
above, there are two choices:
• scoring with both tables:
[mapping]
0 T 0
0 T 1
• scoring with either table:
[mapping]
0 T 0
1 T 1
Note: what we are really doing here is using Moses’ capabilities to use different encoding paths.
The number before "T" defines a decoding path, so in the second example are two different
decoding paths specified. Decoding paths may also contain additional mapping steps, such as
generation steps and translation steps using different factors.
170
4. User Guide
Also note that there is no way to have the option "use both tables, if the phrase pair is in both
table, otherwise use only the table where you can find it". Keep in mind, that scoring a phrase
pair involves a cost and lowers the chances that the phrase pair is used. To effectively use this
option, you may create a third table that consists of the intersection of the two phrase tables,
and remove shared phrase pairs from each table.
Backoff Models: You may want to prefer to use the first table, and the second table only if
there are no translations to be found in the first table. In other words, the second table is only
a back-off table for unknown words and phrases in the first table. This can be specified by
the option decoding-graph-back-off. The option also allows if the back-off table should only
be used for single words (unigrams), unigrams and bigrams, everything up to trigrams, up to
4-grams, etc.
For example, if you have two translation tables, and you want to use the second one only for
unknown words, you would specify:
[decoding-graph-backoff]
0
1
The 0 indicates that the first table is used for anything (which it always should be), and the 1
indicates that the second table is used for unknown n-grams up to size 1. Replacing it with a 2
would indicate its use for unknown unigrams and bigrams (unknown in the sense that the first
table has no translations for it).
Also note, that this option works also with more complicated mappings than just a single translation table. For instance the following specifies the use of a simple translation table first, and
as a back-off a more complex factored decomposition involving two translation tables and two
generation tables:
[mapping]
0 T 0
1 T 1
1 G 0
1 T 2
1 G 1
[decoding-graph-backoff]
0
1
4.3.21
Pruning the Translation Table
The translation table contains all phrase pairs found in the parallel corpus, which includes a lot
of noise. To reduce the noise, recent work by Johnson et al. has suggested to prune out unlikely
phrase pairs. For more detail, please refer to the paper:
4.3.
171
H. Johnson, J. Martin, G. Foster and R. Kuhn. (2007) ”’Improving Translation Quality by Discarding Most of the Phrasetable”’. In Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning
(EMNLP-CoNLL), pp. 967-975.
Build Instructions
Moses includes a re-implementation of this method in the directory contrib/sigtest-filter.
You first need to build it from the source files.
This implementation relies on Joy Zhang’s. The source code can be download from github46 .
Joy’s original code is here47 .
1.
2.
3.
4.
in SALM/Distribution/Linux type: make
enter the directory contrib/sigtest-filter in the main Moses distribution directory
type make SALMDIR=/path/to/SALM
Usage Instructions
Using the SALM/Bin/Linux/Index/IndexSA.O32, create a suffix array index of the source and
target sides of your training bitext (SOURCE, TARGET).
% SALM/Bin/Linux/Index/IndexSA.O32 TARGET
% SALM/Bin/Linux/Index/IndexSA.O32 SOURCE
Prune the phrase table:
% cat phrase-table | ./filter-pt -e TARGET -f SOURCE -l FILTER-VALUE > phrase-table.pruned
FILTER-VALUE is the -log prob threshold described in Johnson et al. (2007)’s paper. It may be
either ’a+e’, ’a-e’, or a positive real value. Run with no options to see more use-cases. A
good setting is -l a+e -n 30, which also keeps only the top 30 phrase translations for each
source phrase, based on p(e|f).
If you filter an hierarchical model, add the switch -h.
46
47
https://github.com/moses-smt/salm
http://projectile.sv.cmu.edu/research/public/tools/salm/salm.htm#update
172
4. User Guide
Using the EMS
To use this method in experiment.perl, you will have to add two settings in the TRAINING
section:
salm-index = /path/to/project/salm/Bin/Linux/Index/IndexSA.O64
sigtest-filter = "-l a+e -n 50"
The setting salm-index points to the binary to build the suffix array, and sigtest-filter
contains the options for filtering (excluding -e, -f, -h). EMS detects automatically, if you filter
a phrase-based or hierarchical model and if a reordering model is used.
4.3.22
Pruning the Phrase Table based on Relative Entropy
While the pruning method in Johnson et al. (2007)48 is designed to remove spurious phrase
pairs due to noisy data, it is also possible to remove phrase pairs that are redundant. That is,
phrase pairs that can be composed by smaller phrase pairs in the model with similar probabilities. For more detail please refer to the following papers:
Ling, W., Graça, J., Trancoso, I., and Black, A. (2012). Entropy-based Pruning for Phrase-based
Machine Translation. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL),
pp. 962-971.
Zens, R., Stanton, D., Xu, P. (2012). A Systematic Comparison of Phrase Table Pruning Technique. In Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language
Processing and Computational Natural Language Learning (EMNLP-CoNLL), pp. 972-983.
The code from Ling et al. (2012)’s paper is available at contrib/relent-filter.
Update The code in contrib/relent-filter no longer works with the current version of Moses. To
compile it, use an older version of Moses with this command:
git checkout RELEASE-0.91
Build Instructions
The binaries for Relative Entropy-based Pruning are built automatically with Moses. However, this implementation also calculates the significance scores (Johnson et al., 2007)49 , using a
slightly modified version of the code by Chris Dyer, which is in contrib/relent-filter/sigtest-filter.
This must be built using the same procedure:
48
49
http://www.aclweb.org/anthology/D/D07/D07-1103.pdf
http://www.aclweb.org/anthology/D/D07/D07-1103.pdf
4.3.
173
2. Run "make SALMDIR=/path/to/SALM" in "contrib/relent-filter/sigtest-filter" to
create the executable filter-pt
Usage Instructions
Checklist of required files (I will use <varname> to refer to these vars):
1.
2.
3.
4.
s_train - source training file
t_train - target training file
moses_ini - path to the Moses configuration file ( after tuning )
pruning_binaries - path to the relent pruning binaries ( should be "bin" if no changes
5. pruning_scripts - path to the relent pruning scripts ( should be "contrib/relent-filter/scripts"
if no changes were made )
6. sigbin - path to the sigtest filter binaries ( should be "contrib/relent-filter/sigtest-filter"
if no changes were made )
7. output_dir - path to write the output
Build suffix arrays for the source and target parallel training data
% SALM/Bin/Linux/Index/IndexSA.O32 <s_train>
% SALM/Bin/Linux/Index/IndexSA.O32 <t_train>
Calculate phrase pair scores by running:
% perl <pruning_scripts>/calcPruningScores.pl -moses_ini <moses_ini> \
-training_s <s_train> -training_t <t_train> \
-prune_bin <pruning_binaries> -prune_scripts <pruning_scripts> \
-moses_scripts <path_to_moses>/scripts/training/ \
-workdir <output_dir> -dec_size 10000
This will create the following files in the <output_dir>/scores/ dir:
1.
2.
3.
4.
5.
50
count.txt - counts of the phrase pairs for N(s,t) N(s,*) and N(*,t)
divergence.txt - negative log of the divergence of the phrase pair
empirical.txt - empirical distribution of the phrase pairs N(s,t)/N(*,*)
rel_ent.txt - relative entropy of the phrase pairs
significance.txt - significance of the phrase pairs
http://projectile.sv.cmu.edu/research/public/tools/salm/salm.htm#update
174
4. User Guide
You can use any one of these files for pruning and also combine these scores using the script
<pruning_scripts>/interpolateScores.pl.
To actually prune a phrase table, run <pruning_scripts>/prunePT.pl, this will prune phrase
pairs based on the score file that is used. This script will prune the phrase pairs with lower
scores first.
For instance, to prune 30% of the phrase table using relative entropy run:
% perl <pruning_scripts>/prunePT.pl -table <phrase_table_file> \
-scores <output_dir>/scores/rel_ent.txt -percentage 70 > <pruned_phrase_table_file>
You can also prune by threshold
% perl <pruning_scripts>/prunePT.pl -table <phrase_table_file> \
-scores <output_dir>/scores/rel_ent.txt -threshold 0.1 > <pruned_phrase_table_file>
The same must be done for the reordering table by replacing <phrase_table_file> with the
<reord_table_file>
% perl <pruning_scripts>/prunePT.pl -table <reord_table_file> \
-scores <output_dir>/scores/rel_ent.txt -percentage 70 > <pruned_reord_table_file>
Parallelization
The script <pruning_scripts>/calcPruningScores.pl requires the forced decoding of the
whole set of phrase pairs in the phrase table, so unless it is used for a small corpora, it usually
requires large amounts of time to process. Thus, we recommend users to run multiple instances
of <pruning_scripts>/calcPruningScores.pl in parallel to process different parts of the
phrase table.
To do this, run:
% perl <pruning_scripts>/calcPruningScores.pl -moses_ini <moses_ini> \
-training_s <s_train> -training_t <t_train> \
-prune_bin <pruning_binaries> -prune_scripts <pruning_scripts> \
-moses_scripts <path_to_moses>/scripts/training/ \
-workdir <output_dir> -dec_size 10000 -start 0 -end 100000
4.3.
175
The -start and -end options tell the script to only calculate the results for phrase pairs between
0 and 99999.
Thus, an example of a shell script to run for the whole phrase table would be:
size=‘wc <phrase_table_file> | gawk ’{print $1}’‘ phrases_per_process=100000 for i in$(seq 0 $phrases_per_process$size)
do
end=‘expr $i +$phrases_per_process‘
perl <pruning_scripts>/calcPruningScores.pl -moses_ini <moses_ini> \
-training_s <s_train> -training_t <t_train> \
-prune_bin <pruning_binaries> -prune_scripts <pruning_scripts> \
-moses_scripts <path_to_moses>/scripts/training/
-workdir <output_dir>.$i-$end -dec_size 10000 -start $i -end$end
done
After all processes finish, simply join the partial score files together in the same order.
4.3.23
Moses supports multi-threaded operation, enabling faster decoding on multi-core machines.
The current limitations of multi-threaded Moses are:
1. irstlm is not supported, since it uses a non-threadsafe cache
2. lattice input may not work - this has not been tested
3. increasing the verbosity of Moses will probably cause multi-threaded Moses to crash
Multi-threaded Moses is now built by default. If you omit the -threads argument, then Moses
will use a single worker thread, and a thread to read the input stream. Using the argument
-threads n specifies a pool of n threads, and -threads all will use all the cores on the machine.
4.3.24
Moses Server
The Moses server enables you to run the decoder as a server process, and send it sentences to
be translated via XMLRPC51 . This means that one Moses process can service distributed clients
coded in Java, perl, python, php, or any of the many other languages which have XMLRPC
libraries.
To build the Moses server, you need to have XMLRPC-c52 installed and you need to add the argument --with-xmlrpc-c=<path-xmlrpc-c-config> to the configure arguments. It has been
51
52
http://www.xmlrpc.com/
http://xmlrpc-c.sourceforge.net/
176
4. User Guide
tested with the latest stable version, 1.16.19. You will also need to configure Moses for multithreaded operation, as described above.
Running make should then build an executable server/mosesserver. This can be launched
using the same command-line arguments as moses, with two additional arguments to specify
the listening port and log-file (--server-port and --server-log). These default to 8080 and
/dev/null respectively.
A sample client is included in the server directory (in perl), which requires the SOAP::Lite
perl module installed. To access the Moses server, an XMLRPC request should be sent to
http://host:port/RPC2 where the parameter is a map containing the keys text and (optionally)
align. The value of the first of these parameters is the text to be translated and the second, if
present, causes alignment information to be returned to the client. The client will receive a map
containing the same two keys, where the value associated with the text key is the translated
text, and the align key (if present) maps to a list of maps. The alignment gives the segmentation in target order, with each list element specifying the target start position (tgt-start),
source start position (src-start) and source end position (src-end).
Note that although the Moses server needs to be built against multi-threaded moses, it can
be run in single-threaded mode using the --serial option. This enables it to be used with
non-threadsafe libraries such as (currently) irstlm.
Using Multiple Translation Systems in the Same Server
Alert: This functionality has been removed as of May 2013. A replacement is Alternate
Weight Settings53 .
The Moses server is now able to load multiple translation systems within the same server,
and the client is able to decide which translation system that the server should use, on a persentence basis. The client does this by passing a system argument in the translate operation.
One possible use-case for this multiple models feature is if you want to build a server that
translates both French and German into English, and uses a large English language model.
Instead of running two copies of the Moses server, each with a copy of the English language
model in memory, you can now run one Moses server instance, with the language model in
memory, thus saving on RAM.
To use the multiple models feature, you need to make some changes to the standard Moses
configuration file. A sample configuration file can be found here54 .
The first piece of extra configuration required for a multiple models setup is to specify the
available systems, for example
[translation-systems]
de D 0 R 0 L 0
fr D 1 R 1 L 1
53
54
http://www.statmt.org/moses/img/moses-en.ini
4.3.
177
This specifies that there are two systems (de and fr), and that the first uses decode path 0,
reordering model 0, and language model 0, whilst the second uses the models with id 1. The
multiple decode paths are specified with a stanza like
[mapping]
0 T 0
1 T 1
which indicates that the 0th decode path uses the 0th translation model, and the 1st decode
path uses the 1st translation model. Using a language model specification like
[lmodel-file]
0 0 5 /disk4/translation-server/models/interpolated-lm
0 0 5 /disk4/translation-server/models/interpolated-lm
means that the same language model can be used in two different systems with two different
weights, but Moses will only load it once. The weights sections of the configuration file must
have the correct numbers of weights for each of the models, and there must be a word penalty
and linear distortion weight for each translation system. The lexicalised reordering weights (if
any) must be specified in the [weight-lr] stanza, with the distortion penalty in the [weight-d]
stanza.
4.3.25
Continue Partial Translation
Alert: This functionality has been removed as of May 2013.
This option forces Moses to start generating the translation from a non-empty hypothesis. This
can be useful in situations, when you have already translated some part of the sentence and
want to get a suggestion or an n-best-list of continuations.
Use -continue-partial-translation (-cpt) to activate this feature. With -cpt, Moses accepts
also a special format of the input: three parameters delimited by the triple bar (|||). The first
parameter is the string of output produced so far (used for LM scoring). The second parameter
is the coverage vector of input words are already translated by the output so far, written as a
string of "1"s and "0"s of the same length as there are words in the input sentence. The third
parameter is the source sentence.
Example:
% echo "that is ||| 11000 ||| das ist ein kleines haus" | moses -f moses.ini -continue-partial-translation
that is a small house
% echo "that house ||| 10001 ||| das ist ein kleines haus" | moses -f moses.ini -continue-partial-translation
that house is a little
178
4. User Guide
If the input does not fit to this pattern, it is treated like normal input with no words translated
yet.
This type of input is currently not compatible with factored models or confusion networks. The
standard non-lexicalized distortion works but more or less as one would expect (note that some
input coverage vectors may prohibit translation under low distortion limits). The lexicalized
reordering has not been tested.
Options
• -continue-partial-translation (-cpt) -- activate the feature
4.3.26
Global Lexicon Model
The global lexicon model predicts the bag of output words from the bag of input words. It
does not use an explicit alignment between input and output words, so word choice is also
influenced by the input context. For details, please check Mauser et al., (2009)55 .
The model is trained with the script
scripts/training/train-global-lexicon-model.perl --corpus-stem FILESTEM --lex-dir DIR --f EXT --e EXT
which requires the tokenized parallel corpus, and the lexicon files required for GIZA++.
You will need the MegaM56 maximum entropy classifier from Hal Daume for training.
Warning: A separate maximum entropy classifier is trained for each target word, which is
very time consuming. The training code is a very experimental state. It is very inefficient. For
instance training a model on Europarl German-English with 86,700 distinct English words took
The model is stored in a text file.
File format:
county
county
county
county
county
county
initiativen 0.34478
land 0.92405
schaffen 0.23749
stehen 0.39572
weiteren 0.04581
europa -0.47688
Specification in moses.ini:
55
56
http://www.cs.utah.edu/~hal/megam/
4.3.
179
[feature]
GlobalLexicalModel input-factor=0 output-factor=0 path=.../global-lexicon.gz
[weight]
GlobalLexicalModel0= 0.1
4.3.27
Incremental Training
Introduction
Translation models for Moses are typically batch trained. That is, before training you have all
the data you wish to use, you compute the alignments using GIZA, and from that produce a
phrase table which you can use in the decoder. If some time later you wish to utilize some new
training data, you must repeat the process from the start, and for large data sets, that can take
quite some time.
Incremental training provides a way of avoiding having to retrain the model from scratch every time you wish to use some new training data. Instead of producing a phrase table with
precalculated scores for all translations, the entire source and target corpora are stored in memory as a suffix array along with their alignments, and translation scores are calculated on the
fly. Now, when you have new data, you simply update the word alignments, and append the
new sentences to the corpora along with their alignments. Moses provides a means of doing
this via XML RPC, so you don’t even need to restart the decoder to use the new data.
Note that at the moment the incremental phrase table code is not thread safe.
Initial Training
This section describes how to initially train and use a model which support incremental training.
• Setup the EMS as normal, but use a modified version of GIZA++57 .
training-options = "-final-alignment-model hmm"
to the TRAINING section of your experiment configuration file.
• Train the system using the initial training data as normal.
• See the section below, How to use memory-mapped dynamic suffix array phrase tables
in the moses decoder58 .
57
58
180
4. User Guide
DEPRECATED:
<del> * Modify the moses.ini file found in <experiment-dir> /evaluation/filtered.<evaluation-set>.<
to have a ttable-file entry as follows:</del>
<del>PhraseDictionaryDynSuffixArray source=<path-to-source-corpus> target=<path-to-targetcorpus> alignment=<path-to-alignments></del>
<del>The source and target corpus paths should be to the tokenized, cleaned, and truecased
versions found in <experiment-dir>/training/corpus.<run>.<lang>, and the alignment
path should be to <experiment-dir>/model/aligned.<run>.grow-diag-final-and.</del>
How to use memory-mapped dynamic suffix array phrase tables in the moses decoder
(phrase-based decoding only)
1. Compile with the bjam switch --with-mm
2. You need
- sentences aligned text files
- the word alignment between these files in symal output format
3. Build binary files
Let
${L1} be the extension of the language that you are translating from,${L2} the extension of the language that you want to translate into, and
{CORPUS} the name of the word-aligned training corpus % % % % zcat{CORPUS}.${L1}.gz | mtt-build -i -o /some/path/${CORPUS}.${L1} zcat${CORPUS}.${L2}.gz | mtt-build -i -o /some/path/${CORPUS}.${L2} zcat${CORPUS}.${L1}-${L2}.symal.gz | symal2mam /some/path/${CORPUS}.${L1}-${L2}.mam mmlex-build /some/path/${CORPUS} ${L1}${L2} -o /some/path/${CORPUS}.${L1}-${L2}.lex -c /some/path/${CORPUS}.${L1}-${L2}.coc
4. Define line in moses.ini :
PhraseDictionaryBitextSampling name=PT0 output-factor=0 num-features=9 path=/some/path/${CORPUS} L1=${L1} L2=${L2} pfwd=g pbwd=g smooth=0 sample=1000 workers=1 You can increase the number of workers for sampling (a bit faster), but you’ll lose replicability of the translation output. (The best configuration of phrase table features is still under investigation.) 4.3. Advanced Features of the Decoder 181 Updates Preprocess New Data First, tokenise, clean, and truecase both target and source sentences (in that order) in the same manner as for the original corpus. You can see how this was done by looking at the <experiment-dir>/steps/<run>/CORPUS_{tokenize,clean,truecase}.<run> scripts. Prepare New Data The preprocessed data now needs to be prepared for use by GIZA. This involves updating the vocab files for the corpus, converting the sentences into GIZA’s snt format, and updating the cooccurrence file. plain2snt$ $INC_GIZA_PP/GIZA++-v2/plain2snt.out <new-source-sentences> <new-target-sentences> \ -txt1-vocab <previous-source-vocab> -txt2-vocab <previous-target-vocab> The previous vocabulary files for the original corpus can be found in <experiment-dir>/training/prepared.<r Running this command with the files containing your new tokenized, cleaned, and truecased source and target as txt1 and txt2 will produce new a new vocab file for each language and a couple of .snt files. Any further references to vocabs in commands or config files should reference the new vocabulary files just produced. Note: if this command fails with the error message plain2snt.cpp:28: int loadVocab(): Assertion ‘iid1.size()-1 == ID’ failed., then change line 15 in plain2snt.cpp to vector<string> iid1(1),iid2(1); and recompile. snt2cooc$ $INC_GIZA_PP/bin/snt2cooc.out <new-source-vcb> <new-target-vcb> <new-source_target.snt> \ <previous-source-target.cooc > new.source-target.cooc$ INC_GIZA_PP/bin/snt2cooc.out <new-target-vcb> <new-source-vcb> <new-target_source.snt> \ <previous-target-source.cooc > new.target-source.cooc This commands is run once in the source-target direction, and once in the target-source direction. The previous cooccurrence files can be found in <experiment-dir>/training/giza.<run>/<target-lan and <experiment-dir>/training/giza-inverse.<run>/<source-lang>-<target-lang>.cooc. Update and Compute Alignments GIZA++ can now be run to update and compute the alignments for the new data. This should be run in the source to target, and target to source directions. A sample GIZA++ config file is given below for the source to target direction; for the target to source direction, simply swap mentions of target and source. 182 4. User Guide S: <path-to-src-vocab> T: <path-to-tgt-vocab> C: <path-to-src-to-tgt-snt> O: <prefix-of-output-files> coocurrencefile: <path-to-src-tgt-cooc-file> model1iterations: 1 model1dumpfrequency: 1 hmmiterations: 1 hmmdumpfrequency: 1 model2iterations: 0 model3iterations: 0 model4iterations: 0 model5iterations: 0 emAlignmentDependencies: 1 step_k: 1 oldTrPrbs: <path-to-original-thmm> oldAlPrbs: <path-to-original-hhmm> To run GIZA++ with these config files, just issue the command GIZA++ <path-to-config-file> With the alignments updated, we can get the alignments for the new data by running the command: giza2bal.pl -d <path-to-updated-tgt-to-src-ahmm> -i <path-to-updated-src-to-tgt-ahmm> \ | symal -alignment="grow" -diagonal="yes" -final="yes" -both="yes" > new-alignment-file • Update Model Now that alignments have been computed for the new sentences, you can use them in the decoder. Updating a running Moses instance is done via XML RPC, however to make the changes permanent, you must append the tokenized, cleaned, and truecased source and target sentences to the original corpora, and the new alignments to the alignment file. 4.3.28 Distributed Language Model NB - THIS HAS BEEN REMOVED FROM MOSES (HIEU) In most cases, MT output improves significantly when more data is used to train the Language Model. More data however produces larger models, and it is very easy to produce a model 4.3. Advanced Features of the Decoder 183 which cannot be held in the main memory of a single machine. To overcome this, the Language Model can be distributed across many machines, allowing more data to be used at the cost of a performance overhead. Support for Distributed Language Models in Moses are built on top of a bespoke distributed map implementation called DMap. DMap and support for Distributed Language Models are still in beta, and any feedback or bug reports are welcomed. Installing and Compiling Before compiling Moses with DMap support, you must configure your DMap setup (see below). Once that has been done, run Moses’ configure script with your normal options and --with-dmaplm=<path-to-dmap>, then the usual make, make install. Configuration Configuring DMap is at the moment, a very crude process. One must edit the src/DMap/Config.cpp file by hand and recompile when making any changes. With the configuration being compiled in, this also means that once it is changed, any programs statically linked to DMap will have to be recompiled too. The file src/DMap/Config.cpp provides a good example configuration which is self explanatory. Example In this example scenario, we have a Language Model trained on the giga4 corpus which we wish to host across 4 servers using DMap. The model is a 5-gram model, containing roughly 210 million ngrams; the probabilities and backoff weights of ngrams will be uniformly quantised to 5 bit values. Configuration Here is an example Config.cpp for such a set up: config->setShardDirectory("/home/user/dmap"); config->addTableConfig(new TableConfigLossyDoubleHash( "giga4", // name of table 283845991, // number of cells (approx 1.23 * number of ngrams) 64, // number of chunks (not too important, leave at 64) (((uint64_t)1 << 61) - 1), // universal hashing P parameter 5789372245 % (((uint64_t)1 << 61) - 1), // universal hashing a parameter 3987420741 % (((uint64_t)1 << 61) - 1), // universal hashing b parameter "/home/user/dmap/giga4.bf", 16, // num_error_bits (higher -> fewer collisions but more memory) 10, // num_value_bits (higher -> more accurate probabilities // and backoff weights but more memory) 20)); // num_hashes (higher -> 184 4. User Guide // config->addStructConfig(new StructConfigLanguageModelBackoff( "giga4", // struct name "giga4", // lm table name 5, // lm order 5, // num logprob bits (these fields should add up to the number // of value bits for the table) 5)); // num backoff bits config->addServerConfig(new ServerConfig("server0.some.domain", 5000)); config->addServerConfig(new ServerConfig("server1.some.domain", 5000)); config->addServerConfig(new ServerConfig("server2.some.domain", 5000)); config->addServerConfig(new ServerConfig("server3.some.domain", 5000)); Note that the shard directory should be on a shared file system all Servers can access. Create Table The command: create_table giga4 will create the files for the shards. Shard Model The model can now be split into chunks using the shard utility: shard giga4 /home/user/dmap/giga4.arpa Create Bloom Filter A Bloom filter is a probabilistic data structure encoding set membership in an extremely space efficient manner. When querying whether a given item is present in the set they encode, they can produce an error with a calculable probability. This error is one sided in that they can produce false positives, but never false negatives. To avoid making slow network requests, DMap keeps a local Bloom filter containing the set of ngrams in the Language Model. Before making a network request to get the probability of an ngram, DMap first checks to see if the ngram is present in the Bloom filter. If is not, then we know for certain the ngram is not present in the model and therefore not worth issuing a network request for. However, if the ngram is present in the filter, it might actually be in the model, or the filter may have produced a false positive. To create a Bloom filter containing the ngrams of the Language Model, run this command: 4.3. Advanced Features of the Decoder 185 ngrams < /home/user/dmap/giga4.arpa | mkbf 134217728 210000000 /home/user/dmap/giga4.bf Integration with Moses The structure within DMap Moses should use as the Language Model should be put into a file, in this case at /home/user/dmap/giga4.conf: giga4 false Note that if for testing or experimentation purposes you would like to have the whole model on the local machine instead of over the network, change the false to true. You must have sufficient memory to host the whole model, but decoding will be significantly faster. To use this, put the following line in your moses.ini file: 11 0 0 5 /home/user/dmap/giga4.conf 4.3.29 Suffix Arrays for Hierarchical Models The phrase-based model uses a suffix array implementation which comes with Moses. If you want to use suffix arrays for hierarchical models, use Adam Lopez’s implementation. The source code for this is currently available in cdec59 . You have to compile cdec so please follow its instructions. You also need to install pycdec cd python python setup.py install Note: the suffix array code requires Python 2.7 or above. If you have Linux installations which are a few years old, check this first. Adam Lopez’s implementation writes the suffix array to binary files, given the parallel training data and word alignment. The Moses toolkit has a wrapper script which simplifies this process: 59 https://github.com/redpony/cdec/ 186 4. User Guide ./scripts/training/wrappers/adam-suffix-array/suffix-array-create.sh \ [path to cdec/python/pkg] \ [source corpus] \ [target corpus] \ [word alignment] \ [output suffix array directory] \ [output glue rules] WARNING - This requires a lot of memory (approximately 10GB for a parallel corpus of 15 million sentence pairs) Once the suffix array has been created, run another Moses wrapper script to extract the translation rules required for a particular set of input sentences. ./scripts/training/wrappers/adam-suffix-array/suffix-array-extract.sh \ [suffix array directory from previous command] \ [input sentences] \ [output rules directory] \ [number of jobs] This command creates one file for each input sentences with just the rules required to decode that sentences. eg. # ls filtered.5/ grammar.0.gz grammar.1.gz grammar.10.gz grammar.3.gz grammar.4.gz grammar.5.gz grammar.7.gz grammar.8.gz grammar.9.gz .... Note - these files are gzipped, and the rules are formatted in the Hiero format, rather than the Moses format. eg. # zcat filtered.5/grammar.out.0.gz | head -1 [X] ||| monsieur [X,1] ||| mr [X,1] ||| 0.178069829941 2.04532289505 1.8692317009 0.268405526876 0.160579100251 0.0 0.0 ||| 0-0 To use these rules in the decoder, put this into the ini file 4.3. Advanced Features of the Decoder 187 PhraseDictionaryALSuffixArray name=TranslationModel0 table-limit=20 \ num-features=7 path=[path-to-filtered-dir] input-factor=0 output-factor=0 PhraseDictionaryMemory name=TranslationModel1 num-features=1 \ path=[path-to-glue-grammar] input-factor=0 output-factor=0 Using the EMS Adam Lopez’s suffix array implementation is integrated into the EMS, where all of the above commands are executed for you. Add the following line to your EMS config file: [TRAINING] suffix-array = [pycdec package path] # e.g. # suffix-array = /home/github/cdec/python/pkg and the EMS will use the suffix array instead of the usual Moses rule extraction algorithms. You can also have multiple extractors running at once [GENERAL] sa_extractors = 8 WARNING: currently the pycdec simply forks itself N times, therefore this will require N times more memory. Be careful with the interaction with multiple evaluations in parallel in EMS and large suffix arrays. 4.3.30 Fuzzy Match Rule Table for Hierachical Models Another method of extracting rules from parallel data is described in (Koehn, Senellart, 2010-1 AMTA)60 and (Koehn, Senellart, 2010-2 AMTA)61 . To use this extraction method in the decoder, add this to the moses.ini file: [feature] PhraseDictionaryFuzzyMatch source=<source/path> target=<target/path> alignment=<alignment/path> 60 61 http://homepages.inf.ed.ac.uk/pkoehn/publications/amta2010.pdf http://homepages.inf.ed.ac.uk/pkoehn/publications/tm-smt-amta2010.pdf 188 4. User Guide It has not yet been integrated into the EMS. Note: The translation rules generated by this algorith is intended to be used in the chart decoder. It can’t be used in the phrase-based decoder. 4.3.31 Translation Model Combination You can combine several phrase tables by linear interpolation or instance weighting using the script contrib/tmcombine/tmcombine.py, or by fill-up using the script contrib/combineptables/combine-ptables.pl. Linear Interpolation and Instance Weighting Linear interpolation works with any models; for instance weighting, models need to be trained with the option -write-lexical-counts so that all sufficient statistics are available. You can set corpus weights by hand, and instance weighting with uniform weights corresponds to a concatenation of your training corpora (except for differences in word alignment). You can also set weights automatically so that perplexity on a tuning set is minimized. To obtain a tuning set from a parallel tuning corpus, use the Moses training pipeline to automatically extract a list of phrase pairs. The file model/extract.sorted.gz is in the right format. An example call: (this weights test/model1 and test/model2 with instance weighting (-m counts) and test/extract as development set for perplexity minimization, and writes the combined phrase table to test/phrase-table_test5) python tmcombine.py combine_given_tuning_set test/model1 test/model2 \\ -m counts -o test/phrase-table_test5 -r test/extract More information is available in (Sennrich, 2012 EACL)62 and contrib/tmcombine/README.md. Fill-up This combination technique is useful when the relevance of the models is known a priori: typically, when one is trained on in-domain data and the others on out-of-domain data. Fill-up preserves all the entries and scores coming from the first model, and adds entries from the other models only if new. Moreover, a binary feature is added for each additional table to denote the provenance of an entry. These binary features work as scaling factors that can be tuned directly by MERT along with other models’ weights. Fill-up can be applied to both translation and reordering tables. Example call, where ptable0 is the in-domain model: 62 http://www.aclweb.org/anthology/E/E12/E12-1055.pdf 4.3. Advanced Features of the Decoder 189 perl combine-ptables.pl --mode=fillup ptable0 ptable1 ... ptableN > ptable-fillup More information is available in (Bisazza et al., 2011 IWSLT)63 and contrib/combine-ptables/README.md. 4.3.32 Online Translation Model Combination (Multimodel phrase table type) Additionally to the log-linear combination of translation models, Moses supports additional methods to combine multiple translation models into a single virtual model, which is then passed to the decoder. The combination is performed at decoding time. In the config, add a feature PhraseDictionaryMultiModel, which refers to its components as follows: [mapping] 0 T 2 [or whatever the zero-based index of PhraseDictionaryMultiModel is] [feature] PhraseDictionaryMemory tuneable=false num-features=4 input-factor=0 output-factor=0 path=/path/to/model1/phrase-table.gz table-limit=20 PhraseDictionaryMemory tuneable=false num-features=4 input-factor=0 output-factor=0 path=/path/to/model2/phrase-table.gz table-limit=20 PhraseDictionaryMultiModel num-features=4 input-factor=0 output-factor=0 table-limit=20 mode=interpolate lambda=0.2,0.8 components=PhraseDictionaryMemory0,PhraseDictionaryMe [weight] PhraseDictionaryMemory0= 0 0 1 0 PhraseDictionaryMemory1= 0 0 1 0 PhraseDictionaryMultiModel0= 0.2 0.2 0.2 0.2 As component models, PhraseDictionaryMemory, PhraseDictionaryBinary and PhraseDictionaryCompact are supported (you may mix them freely). Set the key tuneable=false for all component models; their weights are only used for table-limit pruning, so we recommend 0 0 1 0 0 (which means p(e|f) is used for pruning). There are two additional valid options for PhraseDictionaryMultiModel, mode and lambda. The only mode supported so far is interpolate, which linearly interpolates all component models, and passes the results to the decoder as if they were coming from a single model. Results are identical to offline interpolation with tmcombine.py and -mode interpolate, except for pruning and rounding differences. The weights for each component model can be configured through the key lambda. The number of weights must be one per model, or one per model per feature. Weights can also be set for each sentence during decoding through mosesserver by passing the parameter lambda. See contrib/server/client_multimodel.py for an example. Sentencelevel weights override those defined in the config. With a running Moses server instance, the weights can also be optimized on a tuning set of phrase pairs, using perplexity minimization. This is done with the XMLRPC method optimize and the parameter phrase_pairs, which is an array of phrase pairs, each phrase pair being 63 http://www.mt-archive.info/IWSLT-2011-Bisazza.pdf 190 4. User Guide an array of two strings. For an example, consult contrib/server/client_multimodel.py. Online optimization depends on the dlib library, and requires Moses to be compiled with the flag --with-dlib=/path/to/dlib. Note that optimization returns a weight vector, but does not affect the running system. To use the optimized weights, either update the moses.ini and restart the server, or pass the optimized weights as a parameter for each sentence. Online Computation of Translation Model Features Based on Sufficient Statistics With default phrase tables, only linear interpolation can be performed online. Moses also supports computing translation probabilities and lexical weights online, based on a (weighted) combination of the sufficient statistics from multiple corpora, i.e. phrase and word (pair) frequencies. As preparation, the training option --write-lexical-counts must be used when training the translation model. Then, use the script scripts/training/create_count_tables.py to convert the phrase tables into phrase tables that store phrase (pair) frequencies as their feature values. scripts/training/create_count_tables.py /path/to/model/phrase-table.gz /path/to/model The format for the translation tables in the moses.ini is similar to that of the Multimodel type, but using the feature type PhraseDictionaryMultiModelCounts and additional parameters to specifify the component models. Four parameters are required: components, target-table, lex-f2e and lex-e2f. The files required for the first two are created by create_count_tables.py, the last two during training of the model with --write-lexical-counts. Binarized/compacted tables are also supported (like for PhraseDictionaryMultiModel). Note that for the target count tables, phrase table filtering needs to be disabled (filterable=false). [mapping] 0 T 4 [or whatever the zero-based index of PhraseDictionaryMultiModelCounts is] [feature] PhraseDictionaryMemory tuneable=false num-features=3 input-factor=0 output-factor=0 path=/path/to/model1/count-table.gz table-limit=20 PhraseDictionaryMemory tuneable=false num-features=3 input-factor=0 output-factor=0 path=/path/to/model2/count-table.gz table-limit=20 PhraseDictionaryMemory tuneable=false filterable=false num-features=1 input-factor=0 output-factor=0 path=/path/to/model1/count-table-target.gz PhraseDictionaryMemory tuneable=false filterable=false num-features=1 input-factor=0 output-factor=0 path=/path/to/model2/count-table-target.gz PhraseDictionaryMultiModelCounts num-features=4 input-factor=0 output-factor=0 table-limit=20 mode=instance_weighting lambda=1.0,10.0 components=PhraseDictionaryMemory0 [weight] PhraseDictionaryMemory0= 1 0 0 PhraseDictionaryMemory1= 1 0 0 PhraseDictionaryMemory2= 1 PhraseDictionaryMemory3= 1 PhraseDictionaryMultiModelCounts0= 0.00402447059454402 0.0685647475075862 0.294089113124688 0.0328320356515851 Setting and optimizing weights is done as for the Multimodel phrase table type, but the supported modes are different. The weights of the component models are only used for table-limit pruning, and the weight 1 0 0, which is pruning by phrase pair frequency, is recommended. The following modes are implemented: 4.3. Advanced Features of the Decoder 191 • instance_weighting: weights are applied to the sufficient statistics (i.e. the phrase (pair) frequencies), not to model probabilities. Results are identical to offline optimization with tmcombine.py and -mode counts, except for pruning and rounding differences. • interpolate: both phrase and word translation probabilities (the latter being used to compute lexical weights) are linearly interpolated. This corresponds to tmcombine.py with -mode interpolate and -recompute-lexweights. 4.3.33 Alternate Weight Settings Note: this functionality currently does not work with multi-threaded decoding. You may want to translate different some sentences with different weight settings than others, due to significant differences in genre, text type, style, or even to have separate settings for headlines and questions. Moses allows you to specify alternate weight settings in the configuration file, e.g.: [alternate-weight-setting] id=strong-lm Distortion0= 0.1 LexicalReordering0= 0.1 0.1 0.1 0.1 0.1 0.1 LM0= 1 WordPenalty0= 0 TranslationModel0= 0.1 0.1 0.1 0.1 0 This example specifies a weight setting with the identifying name strong-lm. When translating a sentence, the default weight setting is used, unless the use of an alternate weight setting is specified with an XML tag: <seg weight-setting="strong-lm">This is a small house .</seg> This functionality also allows for the selective use of feature functions and decoding graphs (unless decomposed factored models are used, a decoding graph corresponds to a translation table). Feature functions can be turned off by adding the parameter ignore-ff to the identifier line (names of feature functions, separated by comma), decoding graphs can be ignored with the parameter ignore-decoding-path (number of decoding paths, separated by comma). Note that with these additional options all the capability of the previously (pre-2013) implemented "Translation Systems" is provided. You can even have one configuration file and one Moses process to translate two different language pairs that share nothing but basic features. See the example below for a complete configuration file with exactly this setup. In this case, the default weight setting is not useful since it mixes translation models and language models from both language pairs. 192 4. User Guide [input-factors] 0 # mapping steps [mapping] 0 T 0 1 T 1 [distortion-limit] 6 # feature functions [feature] Distortion UnknownWordPenalty WordPenalty PhraseDictionaryBinary name=TranslationModel0 num-features=5 \ path=/path/to/french-english/phrase-table output-factor=0 LexicalReordering num-features=6 name=LexicalReordering0 \ type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0 \ path=/path/to/french-english/reordering-table KENLM name=LM0 order=5 factor=0 path=/path/to/french-english/language-model lazyken=0 PhraseDictionaryBinary name=TranslationModel1 num-features=5 \ path=/path/to/german-english/phrase-table output-factor=0 LexicalReordering num-features=6 name=LexicalReordering1 \ type=wbe-msd-bidirectional-fe-allff input-factor=0 output-factor=0 \ path=/path/to/german-english/reordering-table KENLM name=LM1 order=5 factor=0 path=/path/to/german-english/language-model lazyken=0 # core weights - not used [weight] Distortion0= 0 WordPenalty0= 0 TranslationModel0= 0 0 0 0 0 LexicalReordering0= 0 0 0 0 0 0 LM0= 0 TranslationModel1= 0 0 0 0 0 LexicalReordering1= 0 0 0 0 0 0 LM1= 0 [alternate-weight-setting] id=fr ignore-ff=LM1,LexicalReordering1 ignore-decoding-path=1 Distortion0= 0.155 LexicalReordering0= 0.074 -0.008 0.002 0.050 0.033 0.042 LM0= 0.152 WordPenalty0= -0.097 TranslationModel0= 0.098 0.065 -0.003 0.060 0.156 4.3. Advanced Features of the Decoder 193 id=de ignore-ff=LM0,LexicalReordering0 ignore-decoding-path=0 LexicalReordering1= 0.013 -0.012 0.053 0.116 0.006 0.080 Distortion0= 0.171 LM0= 0.136 WordPenalty0= 0.060 TranslationModel1= 0.112 0.160 -0.001 0.067 0.006 With this model, you can translate: <seg weight-setting=de>Hier ist ein kleines Haus .</seg> <seg weight-setting=fr>C’ est une petite maison . </seg> 4.3.34 Open Machine Translation Core (OMTC) - A proposed machine translation system standard A proposed standard for machine translation APIs has been developed as part of the MosesCore64 project (European Commission Grant Number 288487 under the 7th Framework Programme). It is called Open Machine Translation Core (OMTC) and defines a service interface for MT interfaces. This approach allows software engineers to wrap disparate MT back-ends such that they look identical to others no matter which flavour of MT system is being wrapped. This provides a standard protocol for ŞtalkingŤ to MT back-ends. In applications where many MT back-ends are to be used, OMTC allows for easier integration of these back-ends. Even in applications where one MT back-end is used, OMTC provides highly cohesive, yet low coupled, interfaces that should allow the back-end to be replaced by another with little effort. OMTC standardises the follow aspects of an MT system: • Resources: A resource is an object that is provided or constructed by a user action for use in an MT system. Examples of resources are: translation memory, glossary, MT engine, or a document. Two base resource types are defined, from which all other resource types are derived, they are primary and derived resources. Primary resources are resource which are constructed outside of the MT system and are made available to it, e.g., through an upload action. Primary resources are used to defined mono- and multi-lingual resources, translation memories and glossaries. Derived resources, on the other hand, are ones which have been constructed by user action inside of the MT system, e.g., a SMT engine. • Sessions: A session is the period of time in which a user interacts with the MT system. The session interface hierarchy supports both user identity and anonymity. Mixin interfaces are, also, defined, to integrate with any authentication system. • Session Negotiation: This is an optional part of the standard and, if used, shall allow a client and the MT server to come to an agreement about which features, resources (this includes exchange and document formats), pre-requisites (e.g. payment) and API version 64 http://www.statmt.org/mosescore/ 194 4. User Guide support is to be expected from both parties. If no agreement can be found then the client’s session should be torn down, but this is completely application defined. • Authorisation: OMTC can integrate with an authorisation system that may be being used in an MT system. It allows users and roles to be mapped into the API. • Machine Translation Engines: Machine translation engines are derived resources which are capable of performing machine translation of, possibly, unseen sentences. An engine may be an SMT decoding pipeline, for instance. It is application defined as to how this part of the API is implemented. Optionally engine functionality can be mixed-in in order to add the following operations: composition, evaluation, parameter updating, querying, (re-)training, testing and updating. Potentially long running tasks return tickets in order for the application to track these tasks. • Translators: Translators, as defined in OMTC, are a derived resource and are a conglomeration of, at least one of the following, an MT engine, a collection of translation memories, and a collection of glossaries. The translator interface provides methods for translation with returned tickets due to the long running nature of these tasks. A reference implementation of OMTC has been constructed in Java v1.7. It is available in the contrib/omtc directory of the mosesdecoder as a Git submodule. Please see the contrib/omtc/README for details. 4.3.35 Pipeline Creation Language (PCL) Building pipelines can be tedious and error-prone. Using Moses scripts to build pipelines can be hampered by the fact that scripts need to be able to parse the output of the previous script. Moving scripts to different positions in the pipeline is tricky and may require a code change! It would be better if the scripts were re-usable without change and users can start to build up a library of computational pieces that can be used in any pipeline in any position. Since pipelines are widely used in machine translation, and given the problem outlined above, a more convienent and less error prone way of building pipelines quickly, with re-usable components, would aid construction. A domain specific language called Pipeline Creation Language (PCL) has been developed part of the MosesCore65 project (European Commission Grant Number 288487 under the 7th Framework Programme). PCL enables users to gather components into libraries, or packages, and re-use them in pipelines. Each component defines inputs and outputs which are checked by the PCL compiler to verify components are compatible with each other. PCL is a general purpose language that can be used to construct non-recurrent software pipelines. In order to adapt your existing programs and script for use with PCL a Python wrapper must be defined for each program. This builds up a library of components with are combined with others in PCL files. The Python wrapper scripts must implement the following function interface: • get_name() - Returns an object representing the name of the component. The __str__() function should be implemented to return a meaningful name. • get_inputs() - Returns the inputs of the component. Components should only be defined with one input port. A list of input names must be returned. 65 http://www.statmt.org/mosescore/ 4.3. Advanced Features of the Decoder 195 • get_outputs() - Returns the outputs of the component. Components should only be defined with one output port. A list of output names must be returned. • get_configuration() - Returns a list of names that represent the static data that shall be used to construct the component. • configure(args) - This function is the component designer’s chance to preprocess configuration injected at runtime. The args parameter is a dictionary that contains all the configuration provided to the pipeline. This function is to filter out, and optionally preprocess, the configuration used by this component. This function shall return an object containing the configuration necessary to construct this component. • initialise(config) - This function is where the component designer defines the component’s computation. The function receives the output object from the configure() function and must return a function that takes two parameters, an input object, and a state object. The input object is a dictionary that is received from the previous component in the pipeline, and the state object is the configuration for the component. The returned function should be used to define the component’s computation. Once your library of components have been written they can be combined using the PCL language. A PCL file defines one component which uses other defined components. For example, the following file defines a component that performs tokenisation for source and target files. # # Component definition: 2 input ports, 2 output ports # # +---------+ # src_filename -->+ +--> tokenised_src_filename # | | # trg_filename -->+ +--> tokenised_trg_filename # +---------+ # import wrappers.tokenizer.tokenizer as tokeniser component src_trg_tokeniser inputs (src_filename), (trg_filename) outputs (tokenised_src_filename), (tokenised_trg_filename) configuration tokeniser.src.language, tokeniser.src.tokenisation_dir, tokeniser.trg.language, tokeniser.trg.tokenisation_dir, tokeniser.moses.installation declare src_tokeniser := new tokeniser with tokeniser.src.language -> language, tokeniser.src.tokenisation_dir -> tokenisation_dir, tokeniser.moses.installation -> moses_installation_dir trg_tokeniser := new tokeniser with tokeniser.trg.language -> language, tokeniser.trg.tokenisation_dir -> tokenisation_dir, tokeniser.moses.installation -> moses_installation_dir 196 4. User Guide as wire (src_filename -> filename), (trg_filename -> filename) >>> (src_tokeniser *** trg_tokeniser) >>> wire (tokenised_filename -> tokenised_src_filename), (tokenised_filename -> tokenised_trg_filename) A PCL file is composed of the following bits: • Imports: Optional imports can be specified. Notice that all components must be given an alias, in this case the component wrappers.tokenizer.tokenizer shall be referenced in this file by the name tokeniser. • Component: This starts the component definition and provides the name. The component’s name must be the same as the filename. E.g., a component in fred.pcl must be called fred. • Inputs: Defines the inputs of the component. The example above defines a component with a two port input. Specifing a comma-separated list of names defines a one port input. • Outputs: Defines the outputs of the component. The example above defines a component with a two port output. Specifing a comma-separated list of names defines a one port output. • Configuration: Optional configuration for the component. This is static data that shall be used to construct components used in this component. • Declarations: Optional declarations of components used in this component. Configuration is used to construct imported components • Definition: The as portion of the component definition is an expression which defines how the construct components are to be combined to create the computation required for the component. The definition of a component can use the following pre-defined components: • first - This component takes one expression with a one port input and creates a two port input and output component. The provided component is applied only to the first port of the input. • second - This component takes one expression with a one port input and creates a two port input and output component. The provided component is applied only to the second port of the input. • split - Split is a component with one input port and two output ports. The value of the outputs is the input, i.e., spliting the input. • merge - Merge values from the two port input to a one port output. A comma-separated list of top and bottom keywords subscripted with input names are used to map these values to a new name. E.g., merge top[a] -> top_a, bottom[b] -> bottom_b takes the a value of the top input and maps that value to a new name top_a, and the b value of the bottom input and maps that value to a new name bottom_b. • wire - Wires are used to adapt one component’s output to another’s input. For wires with one input and output port then the wire mapping is a comma-separated mapping, e.g., wire a -> next_a, b -> next_b adapts a one port output component whose outputs are a and b to a one port component whose inputs are next_a and next_b. For wires with 4.3. Advanced Features of the Decoder 197 two input and output ports mappings are in comma-separated parenthese, e.g., wire (a -> next_a, b -> next_b), (c -> next_c, d -> next_d). This wire adapts the top input from a to next_a, and b to next_b, and the bottom input from c to next_c and d to next_d. • if - Conditional execution of a component can be achieved with the if component. This component takes three arguments: a conditional expression, a then component and an else component. If the condition is evaluated to a truthy value the then component is executed, otherwise the else component is executed. See the conditional example in the PCL Git repository for an example of usage. Combinator operators used to compose the pipeline, they are: • >>> - Composition. This operator composes two components. E.g., a >>> b creates a component in which a is executed before b. • *** - Parallel execution. This operator creates a component in which the two components provided are executed in parallel. E.g., a *** b creates a component with two input and output ports. • &&& - Parallel execution. The operator creates a component in which two components are executed in parallel from a single input port. E.g., a &&& b creates a component with one input port and two output ports. Examples in the PCL Git repository show the usage of these operators and pre-defined components. Plus an example Moses training pipeline is available in contrib/arrow-pipelines directory of the mosesdecoder Git repository. Please see contrib/arrow-pipelines/README for details of how to compile and run this pipeline. For more details of how to use PCL please see the latest manual at contrib/arrow-pipelines/python/pcl/documentation/pcl-manual.latest.pdf 4.3.36 Placeholder Placeholders are symbols that replaces a word or phrase. For example, numbers (’42.85’) can be replaced with a symbol ’@[email protected]’. Other words and phrases that can potentially be replaced with placeholders include dates and time, and named-entities. This is good in training since the sparse numbers are replaced with more numerous placeholders symbols, producing more reliable statistics for the MT models. The same reason also applies during decoding - the raw number is often an unknown symbol in the phrase-table and language models. Unknown symbols are translated as single words, disabling the advantage of phrasal translation. The reordering of unknown symbols can also be unreliable as we don’t have statistics for it. However, 2 issues arises using placeholder: 1. Translate the original word or phrase. In the example, ’42.85’ should be translated. If the language pair is en-fr, then it may be translated as ’42,85’. 2. How do we insert this translation into the output if the word has be replaced with the placeholder. 198 4. User Guide Moses has support for placeholders in training and decoding. Training When preparing your data, process with data with the script scripts/generic/ph_numbers.perl -c The script was designed to run after tokenization, that is, instead of tokenizing like this: cat [RAW-DATA] | ./scripts/tokenizer/tokenizer.perl -a -l en > TOK-DATA do this cat [RAW-DATA] | ./scripts/tokenizer/tokenizer.perl -a -l en | scripts/generic/ph_numbers.perl -c > TOK-DATA Do this for both source and target language, for parallel and monolingual data. The script will replace numbers with the symbol @[email protected] NB. - this script is currently very simple and language independent. It can be improved to create better translations. During extraction, add the following to the extract command (phrase-based only for now): ./extract --Placeholders @[email protected] .... This will discard any extracted translation rule which are non-consistent with the placeholders. That is, all placeholders must be aligned to 1-to-1 with a placeholder in the other language. Decoding The input sentence must also be processed with the placeholder script to replace numbers with placeholder symbol. However, don’t add the -c argument so that the original number will be retained in the output as an XML entry. For example, 4.3. Advanced Features of the Decoder genericecho
199
"you owe me $100 ." | ./ph_numbers.perl will output you owe me$ <ne translation="@[email protected]" entity="100">@[email protected]</ne> .
Add this to the decoder command when executing the decoder (phrase-based only for now):
./moses
-placeholder-factor 1 -xml-input exclusive
The factor must NOT be one which is being used by the source side of the translation model.
For vanilla models, only factor 0 is used.
The argument -xml-input can be any permitted value, except ’pass-through’.
The output from the decoder will contain the number, not the placeholder. The is the case in
the best output, and the n-best list.
EMS
The above changes can be added to the EMS config file.
For my (Hieu) experiment, these are the changes I made:
1. In the [GENERAL section, change
input-tokenizer = "$misc-script-dir/normalize-punctuation.perl$input-extension | $moses-script-dir/tokenizer/tokenizer.perl -a -l$input-extension"
to
input-tokenizer = "$misc-script-dir/normalize-punctuation.perl$input-extension | $moses-script-dir/tokenizer/tokenizer.perl -a -l$input-extension | $moses-script-dir/gener and change output-tokenizer = "$misc-script-dir/normalize-punctuation.perl $output-extension |$moses-script-dir/tokenizer/tokenizer.perl -a -l $output-extension |$moses-script-dir/ge
to
output-tokenizer = "$misc-script-dir/normalize-punctuation.perl$output-extension | $moses-script-dir/tokenizer/tokenizer.perl -a -l$output-extension | $moses-script-dir/ge 2. In the [TRAINING] section, add extract-settings = "--Placeholders @[email protected]" 3. In the [TUNING] section, change decoder-settings = "-threads 8" to decoder-settings = "-threads 8 -placeholder-factor 1 -xml-input exclusive" And in the [EVALUATION] section, change decoder-settings = "-mbr -mp -search-algorithm 1 -cube-pruning-pop-limit 5000 -s 5000 -threads 8" to decoder-settings = "-mbr -mp -search-algorithm 1 -cube-pruning-pop-limit 5000 -s 5000 -threads 8 -placeholder-factor 1 1 -xml-input exclusive" 4. In the [EVALUATION] section, add input-tokenizer = "$misc-script-dir/normalize-punctuation.perl $input-extension |$moses-script-dir/tokenizer/tokenizer.perl -a -l $input-extension |$moses-script-dir/gener
output-tokenizer = "$misc-script-dir/normalize-punctuation.perl$output-extension | $moses-script-dir/tokenizer/tokenizer.perl -a -l$output-extension
"
200
4. User Guide
Results
This was tested on some experiments, trained with Europarl data. It didn’t have a positive
effect on BLEU score, even reducing it slightly.
However, it may still be helpful to users who translate text with lots of numbers or dates etc.
Also, the recognizer script could be improved.
en-es:
baseline: 24.59.
with placeholder:
es-en:
baseline: 23.00
with placeholder:
en-cs:
baseline: 11.05
with placeholder:
cs-en:
baseline: 15.80
with placeholder:
4.3.37
24.68
22.84
10.62
15.62
Modified Moore-Lewis Filtering
When you have a lot of out-of-domain data and you do not want to use all of it, then you can
filter down that data to the parts that are more similar to the in-domain data. Moses implements
a method called modified Moore-Lewis filtering. The method basically train in-domain and
out-of-domain language models, and removes sentence pairs that receive relatively low scores
by the in-domain models. For more details, please refer to the following paper:
Axelrod, Amittai and He, Xiaodong and Gao, Jianfeng: Domain Adaptation via Pseudo In-Domain
Data Selection, Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing mentioned in Domain Adaptation, pdf66 , bib67 .
The Moses implementation is integrated into EMS. You have to specify in-domain and out-ofdomain in separate CORPUS sections (you can have more than one of each), and then set in the
configuration file which out-of-domain corpora need to be filtered
### filtering some corpora with modified Moore-Lewis
#
mml-filter-corpora = giga
mml-before-wa = "-proportion 0.2"
#mml-after-wa = "-proportion 0.2"
66
67
http://aclweb.org/anthology//D/D11/D11-1033.pdf
http://aclweb.org/anthology//D/D11/D11-1033.bib
4.4.
Sparse Features
201
There are two different places when to do the filtering, either before or after word alignment.
There may be some benefits of having out-of-domain data to improve sentence alignment, but
that may also be computationally to expensive. In the configuration file, you specify the proportion of the out-of-domain data that will be retained - in the example above 20
4.3.38
Constrained Decoding
To constrain the output of the decoder to just the reference sentences, add this as a feature:
[feature]
....
ConstrainedDecoding path=ref.txt
4.4
Sparse Features
Sparse feature functions in Moses allow for thousands of features that follow a specific pattern,
typically lexical instantiations of a general feature function. Take for instance the target word
insertion feature function, which allows the training of lexical indicators for any word (say,
the or fish). Each lexicalized instantiation has its own feature weight, which is typically trained
during tuning. Inserting a the should be fine, inserting the word fish not so much, and the
learned feature weight should reflect this.
In Moses, all feature functions can contain sparse features and dense features. The number of
dense feature has to be specified in advance in moses.ini file, e.g.,
KENLM num-features=1 ...
The decoder doesn’t have to know whether a feature function contains sparse features. And by
definition, the number of sparse features is not specified beforehand.
Sparse lexical features require a special weight file that contains the weight for each instantiation of a feature.
The weight file has to be specified in the moses.ini file:
[weight-file]
path/sparse-weights
202
4. User Guide
This file may look like:
twi_fish -0.5
twi_of -0.001
[...]
By convention, the format for sparse features is
InstanceName_SparseFeatureName
Of course, you want to learn these feature weights during tuning, which requires the use of
either PRO or kbMIRA - it does not work with plain MERT.
4.4.1
Word Translation Features
There are three types of lexical feature function:
• word translation feature, which indicates if a specific source word was translated as a
specific target word
• target word insertion, which indicates if a specific target word has no alignment point
(aligns to no source word in the word alignment stored for the phrase pair)
• source word deletion, which indicates if a specific source word has no alignment point
Specification in moses.ini
The following lines need to be added to the configuration file:
[feature]
TargetWordInsertionFeature factor=FACTOR [path=FILE]
SourceWordDeletionFeature factor=FACTOR [path=FILE]
WordTranslationFeature input-factor=FACTOR output-factor=FACTOR \
[source-path=FILE] [target-path=FILE]-path= \
simple=1 source-context=0 target-context=0
Note that there is no corresponding weight setting for these features.
The optional word list files (one token per line) restrict the feature function to the specified
words. If no word list file is specified, then features for all words a generated.
4.4.
Sparse Features
203
Specification with experiment.perl
Word translation features can be specified as follows:
TRAINING:sparse-features = \
"target-word-insertion top 50, source-word-deletion top 50, \
word-translation top 50 50"
This specifications includes
• target word insertion features for the top 50 most frequent target words
• source word deletion features for the top 50 most frequent source words
• word translation features for the top 50 most frequent target words and top 50 most frequent source words
Instead of top 50, you can also specify all when you do not want to have a restricted word
list.
Moreover, for the word translation feature, by specifying factor 1-2, you can change input
and output factor for the feature. For the deletion and insertion features, there is only one
factor to specify, e.g., factor 1.
4.4.2
Phrase Length Features
The phrase length feature function creates three features for each phrase pair:
• the length of the source phrase (in tokens)
• the length of the target phrase
• the pair of the two values above
For instance, when the phrase ein Riesenhaus is translated into a giant house, then the three
features pl_s2 (2 source words), pl_t3 (3 target words), and pl_2,3 (2 source words into 3
target words) are triggered.
Specification in moses.ini
The following lines need to be added to the configuration file:
[feature]
PhraseLengthFeature
204
4. User Guide
Specification with experiment.perl
The inclusion of the phrase length feature is similar to the word translation feature:
TRAINING:sparse-features = "phrase-length"
In case of using both the phrase length feature and the word translation features, you will need
to include them in the same line.
4.4.3
Domain Features
Domain features flag each phrase pair on in which domain (or more accurately: which subset
of the training data) they occur in.
Specification in moses.ini
Domain features are part of the phrase table, there is no specific support for his particular type
of feature function. A sparse phrase table may include any other arbitrary features. Each line
in the phrase table has to contain an additional field that lists the feature name and its logprobability value.
For example, the following phrase pair contains the domain feature flagging that the phrase
pair occurred in the europarl part of the training corpus:
das Haus ||| the house ||| 0.8 0.5 0.8 0.5 2.718 ||| 0-0 1-1 \
||| 5000 5000 2500 ||| dom_europarl 1
If a phrase table contains sparse features, then this needs to be flagged in the configuration file
by adding the word sparse after the phrase table file name.
Specification with experiment.perl
TRAINING:domain-features = "[sparse ](indicator|ratio|subset)"
There are various settings for domain adaptation features. It requires a domain file that indicates at which lines in the parallel corpus cover lines that stem from different [CORPUS] blocks
(default, when used in experiment.perl, but a different domain-file can be also specified.
These features may included as sparse features or as core features in the phrase table, depending in having the prefix Sparse in the parameter.
There are three kind of features:
4.4.
Sparse Features
205
• Indicator: Each phrase pair is marked if it occurs in a specific domain
• Ratio: Each phrase pair is marked with exp(0) <= log(r) <= exp(1) float feature depending
on the ratio r how often it occurs in corpus r.
• Subset: Similar to the indicator feature, but if a phrase pair occurs in multiple domains if
is marked with these domains in one feature
• Bin (not implemented, the idea is the count bin feature mentioned below but with marking count intervals for each domain).
4.4.4
Count Bin Features
The frequency of a phrase pair in the training data may be a useful to determine its reliability.
The count bin features are integrated into the phrase table, just like the domain features, so
Specification with experiment.perl
The counts of phrase pairs get very sparse for frequent phrases. There are just not that many
phrase pairs that occur exactly 634,343 times. Hence, we bin phrase pairs counts, for instance
phrase pairs that occur once, twice, three to nine times, and more often.
In experiment.perl this is accomplished with an additional switch in score settings. For
the example above this looks like this:
TRAINING:score-settings = "--[Sparse]CountBinFeature 1 2 3 10"
Based on the values that are given, different indicator features are included, depending on
which interval count the phrase pair falls, e.g., ]2;3] = third bin.
4.4.5
Bigram Features
TODO
4.4.6
Soft Matching Features
Models with target syntax require an exact match between nonterminals in a rule and the lefthand-side label of rules that can be substituted into it. With the following rules, a model could
be used to decode ’she slept here’, but not ’she slept on the floor’.
S --> she slept AVP1 ||| sie schlief AVP1
AVP --> here ||| hier
PP --> on the floor ||| auf dem boden
With soft matching, we can allow substitutions of nonterminals even if they do not match.
206
4. User Guide
Specification in moses.ini
The following lines need to be added to the configuration file:
[feature]
SoftMatchingFeature path=FILE
with FILE containing a user-defined list of allowed substitutions. For the example above, the
file needs to contain the following line:
PP AVP
Each substitution (even exact matches) triggers a sparse feature which can be used to prefer
some substitutions over others.
The SoftMatchingFeature operates on the target-side labels and is not (yet) implemented for
the Scope3 and OnDisk phrase tables.
4.5
Translating Web pages with Moses
(Code and documentation written by Herve Saint-Amand.)
We describe a small set of publicly available Perl scripts that provide the mechanisms to translate a Web page by retrieving it, extracting all sentences it contains, stripping them of any font
style markup, translating them using the Moses system, re-inserting them in the document
while preserving the layout of the page, and presenting the result to the user, providing a
seamless translation system comparable to those offered by Google, BabelFish and others.
4.5.1
Introduction
Purpose of this program
Moses is a cutting-edge machine translation program that reflects the latest developments in
the area of statistical machine translation research. It can be trained to translate between any
two languages, and yields high quality results. However, the Moses program taken alone can
only translate plain text, i.e., text stripped of any formatting or style information (as in .txt
files). Also, it only deals with a single sentence at a time.
A program that can translate Web pages is a very useful tool. However, Web pages contain a
lot of formatting information, indicating the color, font and style of each piece of text, along
4.5.
Translating Web pages with Moses
207
with its position in the global layout of the page. Most Web pages also contain more than one
sentence or independent segment of text. For these reasons a Web page cannot be fed directly
to Moses in the hope of obtaining a translated copy.
The scripts described in this document implement a Web page translation system that, at its
core, relies on Moses for the actual translation task. The scripts’ job, given a Web page to
translate, is to locate and extract all strings of text in the page, split paragraphs into individual
sentences, remove and remember any style information associated with the text, send the normalized, plain-text string to Moses for translation, re-apply the style onto the translated text
and re-insert the sentence at its place in the original page structure, (hopefully) resulting in a
translation of the original.
A word of warning
These scripts are a proof-of-concept type of demonstration, and should not be taken for more
than that. They most probably still contain bugs, and possibly even security holes. They are
not appropriate for production environments.
Intended audience and system requirements
This document is meant for testers and system administrators who wish to install and use the
scripts, and/or to understand how they work.
Before starting, the reader should ideally possess basic knowledge of:
•
•
•
•
UNIX-type command-line environments
TCP/IP networking (know what a hostname and a port are)
how to publish a Web page using a CGI script on an Apache server
how to configure and operate the Moses decoder
and have the following resources available:
• an Apache (or similar) Web server
• the possibility of running CPU- and memory-intensive programs, either on the Web server
itself (not recommended), or on one or several other machines that can be reached from
the Web server
• Moses installed on those machines
208
4. User Guide
Overview of the architecture
The following is a quick overview of how the whole system works. An attempt at illustrating
the architecture is in the figure above. File names refer to files available from an Git repository,
1. The Moses system is installed and configured on one or several computers that we designate as Moses servers.
2. On each Moses server, a daemon process, implemented by daemon.pl, accepts network
connections on a given port and copies everything it gets from those connections straight
to Moses, sending back to the client what Moses printed back. This basically plugs Moses
directly onto the network.
3. Another computer, which we designate as the web server, runs Apache (or similar) Web
server software.
4. Through that server, the CGI scripts discussed in this document (index.cgi, translate.cgi
and supporting files) are served to the client, providing the user interface to the system. It
is a simple matter to configure translate.cgi so that it knows where the Moses servers
are located.
5. A client requests index.cgi via the Web server. A form containing a textbox is served
back, where the user can enter a URL to translate.
6. That form is submitted to translate.cgi, which does the bulk of the job. It fetches
the page from the Web, extracts translatable plain text strings from it, sends those to
the Moses servers for translation, inserts the translations back into the document, and
serves the document back to the client. It adjusts links so that if any one is clicked in the
translated document, a translated version will be fetched rather than the document itself.
4.5.
Translating Web pages with Moses
209
The script containing all the interesting code, translate.cgi, is heavily commented, and programmers might be interested in reading it.
Mailing list
Should you encounter problems you can’t solve during the installation and operation of this
program, you can write to the moses support mailing list at [email protected] Should
you not encounter problems, the author (whose email is found in the source file headers) would
be astonished to hear about it.
4.5.2
Detailed setup instructions
Obtaining a copy of the scripts
The scripts are stored in the contrib/web directory in the Moses distribution.
Setting up the Web server
The extracted source code is ready to be run, there is no installation procedure that compiles or
copies files. The program is entirely contained within the directory that was downloaded from
Github. It now needs to be placed on a Web server, in a properly configured location such that
the CGI scripts (the two files bearing the .cgi extension) are executed when requested from a
browser.
For instance, if you are on a shared Web server (e.g., a server provided by your university)
and your user directory contains a directory named public_html, placing the moses-web directory inside public_html should make it available via the Web, at an address similar to
http://www.dept.uni/~you/moses-web/.
Troubleshooting
• 404 Not Found Perhaps the source code folder is not in the right location? Double-check
the directory names. See if the home folder (parent of moses-web itself) is reachable. Ask
• 403 Forbidden, or you see the Perl source code of the script in your browser} The server
is not configured to execute CGI scripts in this directory. Move moses-web to the cgi-bin
subdirectory of your Web home, if it exists. Create a .htacess file in which you enable
the ExecCGI option (see the Apache documentation).
• Internal server error Perhaps the scripts do not have the right permissions to be executed.
Go in moses-web and type the command chmod 755 *cgi.
The scripts are properly installed once you can point your browser at the correct URL and you
see the textbox in which you should enter the URL, and the ’Translate’ button. Pressing the
button won’t work yet, however, as the Moses servers need to be installed and configured first.
210
4. User Guide
Setting up the Moses servers
You now need to install Moses and the daemon.pl script on at least one machine.
Choosing machines for the Moses servers
Running Moses is a slow and expensive process, at least when compared to the world of Web
servers where everything needs to be lightweight, fast and responsive. The machine selected
for running the translator should have a recent, fast processor, and as many GBs of memory as
possible (see the Moses documentation for more details).
Technically, the translations could be computed on the same machine that runs the Web server.
However, the loads that Moses places on a system would risk seriously impacting the performance of the Web server. For that reason, we advise not running Moses on the same computer
as the Web server, especially not if the server is a shared server, where several users host their
files (such as Web servers typically provided by universities). In case of doubt we recommend
For the sake of responsiveness, you may choose to run Moses on several machines at once. The
burden of translation will then be split equally among all the hosts, thus more or less diving
the total translation time by the number of hosts used. If you have several powerful computers
at your disposal, simply repeat the installation instructions that follow on each of the machines
independently.
The main translation script, which runs on the Web server, will want to connect to the Moses
servers via TCP/IP sockets. For this reason, the Moses servers must be reachable from the Web
server, either directly or via SSH tunnels of other proxy mechanisms. Ultimately the translation
script on the Web server must have a hostname/port address it can connect to for each Moses
server.
Installing the scripts
Install Moses For each Moses server, you will need to install and configure Moses for the
language pair that you wish to use. If your Moses servers are all identical in terms of hardware,
OS and available libraries, installing and training Moses on one machine and then copying the
files over to the other ones should work, but your mileage may vary.
Install daemon.pl Once Moses is working, check out, on each Moses server, another copy
of the moses-web source directory by following again the instructions in the download section.
Open bin/daemon.pl, and edit the $MOSES and$MOSES_INI paths to point to the location of
Choose a port number Now you must choose a port number for the daemon process to listen on. Pick any number between 1,024 and 49,151, ideally not a standard port for common
programs and protocols to prevent interference with other programs (i.e., pick a port not mentioned in your /etc/services file).
4.5.
Translating Web pages with Moses
211
Start the daemon To activate a Moses server, simply type, in a shell running on that server:
./daemon.pl <hostname> <port>
where <hostname> is the name of the host you’re typing this on (found by issuing the hostname
command), and <port> is the port you selected. It may be misleading that despite its name,
this program does not fork a background process, it is the background process itself. To truly
launch the process in the background so that it continues running after the shell is closed, this
command might be more useful:
nohup ./daemon.pl <hostname> <port> &
The bin/start-daemon-cluster.pl script distributed with this program provides an automation mechanism that worked well in the original setup on the University of Saarland network.
It was used to start and stop the Moses servers all at once, also setting up SSH tunneling on
startup. Because it is very simple and trimmed to the requirements of that particular installation, we do not explain its use further here, but the reader might find inspiration in reading
it.
Test the Moses servers The daemon should now be listening on the port you chose. When it
receives a connection, it will read the input from that connection one line at a time, passing each
line in turn to Moses for translation, and printing back the translation followed by a newline.
If you have the NetCat tool installed, you can test whether it worked by going to a shell on
the Web server and typing echo "Hello world" | nc <hostname> <port>, replacing Hello
world by a phrase in your source language if it is not English, and <hostname> and <port>
by the values pointing to the Moses server you just set up. A translation should be printed
back.
Configure the tokenizer The translate.cgi script uses external tokenizer and detokenizer
scripts. These scripts adapt their regular expressions depending on the language parsed, and so
tokenizing is improved if the correct language is selected. This is done by opening translate.cgi
with your favourite text editor, and setting $INPUT_LANG and$OUTPUT_LANG to the appropriate
language codes. Currently the existing language codes are the file extensions found in the
bin/nonbreaking_prefixes directory. If yours are not there, simply use en -- end-of-sentence
detection may then be suboptimal, and translation quality may be impacted, but the system
will otherwise still function.
212
4. User Guide
Configure the Web server to connect to the Moses servers The last remaining step is to tell
the frontend Web server where to find the backend Moses servers. Still in translate.cgi, set
the @MOSES_ADDRESSES array to the list of hostname:port strings identifying the Moses servers.
Here is a sample valid configuration for three Moses servers named server01, server02 and
server03, each with the daemon listening on port 7070:
my @MOSES_ADDRESSES = ("server01:7070", "server02:7070", "server03:7070");
Stopping the daemons once done The daemon processes continuously keep a copy of Moses
running, so they consume memory even when idle. For this reason, we recommend that you
stop them once they are not needed anymore, for instance by issuing this command on each
Moses server: killall daemon.pl
5
Training Manual
5.1
Training
5.1.1
Training process
We will start with an overview of the training process. This should give a feel for what is going
on and what files are produced. In the following, we will go into more details of the options of
the training process and additional tools.
The training process takes place in nine steps, all of them executed by the script
train-model.perl
The nine steps are
1.
2.
3.
4.
5.
6.
7.
8.
9.
Prepare data (45 minutes)
Run GIZA++ (16 hours)
Align words (2:30 hours)
Get lexical translation table (30 minutes)
Extract phrases (10 minutes)
Score phrases (1:15 hours)
Build lexicalized reordering model (1 hour)
Build generation models
Create configuration file (1 second)
If you are running on a machine with multiple processors, some of these steps can be considerably sped up with the following option:
--parallel
213
214
5. Training Manual
The run times mentioned in the steps refer to a recent training run on the 751’000 sentence, 16
million word German-English Europarl corpus, on a 3GHz Linux machine.
If you wish to experiment with translation in both directions, step 1 and 2 can be reused, starting from step 3 the contents of the model directory get direction-dependent. In other words run
steps 1 and 2, then make a copy of the whole experiment directory and continue two trainings
from step 3.
5.1.2
Running the training script
For an standard phrase model, you will typically run the training script as follows.
Run the training script:
train-model.perl -root-dir . --corpus corpus/euro --f de --e en
There should be two files in the corpus/ directory called euro.de and euro.en. These files should
be sentence-aligned halfs of the parallel corpus. euro.de should contain the German sentences,
and euro.en should contain the corresponding English sentences.
More on the training parameters (Section 8.3) at the end of this manual. For corpus preparation,
see the section on how to prepare training data (Section 5.2).
5.2
Preparing Training Data
Training data has to be provided sentence aligned (one sentence per line), in two files, one for
the foreign sentences, one for the English sentences:
==> corpus/euro.de <==
wiederaufnahme der sitzungsperiode
ich erklaere die am donnerstag , den 28. maerz 1996 unterbrochene
sitzungsperiode des europaeischen parlaments fuer wiederaufgenommen .
begruessung
==> corpus/euro.en <==
resumption of the session
i declare resumed the session of the european parliament adjourned
on thursday , 28 march 1996 .
welcome
A few other points have to be taken care of:
5.2.
Preparing Training Data
215
• unix commands require the environment variable LC_ALL=C
• one sentence per line, no empty lines
• sentences longer than 100 words (and their corresponding translations) have to be eliminated (note that a shorter sentence length limit will speed up training
• everything lowercased (use lowercase.perl)
5.2.1
Training data for factored models
You will have to provide training data in the format
word0factor0|word0factor1|word0factor2 word1factor0|word1factor1|word1factor2 ...
word0 word1 word2
5.2.2
Cleaning the corpus
The script clean-corpus-n.perl is small script that cleans up a parallel corpus, so it works
well with the training script.
It performs the following steps:
• removes empty lines
• removes redundant space characters
• drops lines (and their corresponding lines), that are empty, too short, too long or violate
the 9-1 sentence ratio limit of GIZA++
The command syntax is:
clean-corpus-n.perl CORPUS L1 L2 OUT MIN MAX
For example: clean-corpus-n.perl raw de en clean 1 50 takes the corpus files raw.de and
raw.en, deletes lines longer than 50, and creates the output files clean.de and clean.en.
216
5. Training Manual
5.3
Factored Training
For training a factored model, you will specify a number of additional training parameters:
--alignment-factors FACTORMAP
--translation-factors FACTORMAPSET
--reordering-factors FACTORMAPSET
--generation-factors FACTORMAPSET
--decoding-steps LIST
Alignment factors
It is usually better to carry out the word alignment (step 2-3 of the training process) on more
general word representations with rich statistics. Even successful word alignment with words
stemmed to 4 characters have been reported. For factored models, this suggests that word
alignment should be done only on either the surface form or the stem/lemma.
Which factors are used during word alignment is set with the --alignment-factors switch.
Let us formally define the parameter syntax:
• FACTOR = [ 0 - 9 ]+
• FACTORLIST = FACTOR [ , FACTOR ]*
• FACTORMAP = FACTORLIST - FACTORLIST
The switch requires a FACTORMAP as argument, for instance 0-0 (using only factor 0 from
source and target language) or 0,1,2-0,1 (using factors 0, 1, and 2 from the source language
and 0 and 1 from the target language).
Typically you may want to train the word alignment using surface forms or lemmas.
5.3.1
Translation factors
The purpose of training factored translation model is to create one or more translation tables
between a subset of the factors. All translation tables are trained from the same word alignment, and are specified with the switch --translation-factors.
To define the syntax, we have to extend our parameter syntax with
• FACTORMAPSET = FACTORMAP[+FACTORMAP]*
since we want to specify multiple mappings.
One example is --translation-factors 0-0+1-1,2, which create the two tables
phrase-table.0-0.gz
phrase-table.1-1,2.gz
5.4.
5.3.2
Training Step 1: Prepare Data
217
Reordering factors
Reordering tables can be trained with --reordering-factors. Syntax is the same as for translation factors.
5.3.3
Generation factors
Finally, we also want to create generation tables between target factors. Which tables to generate is specified with --generation-factors, which takes a FACTORMAPSET as a parameter.
Note that this time the mapping is between target factors, not between source and target factors.
One example is --generation-factors 0-1 with creates a generation table between factor
factor 0 and 1.
5.3.4
Decoding steps
The mapping from source words in factored representation into target words in factored representation takes place in a number of mapping steps (either using a translation table or a
generation table). These steps are specified with the switch --decoding-steps LIST.
For example --decoding-steps t0,g0,t1,t2,g1 specifies that mapping takes place in form of
an initial translation step using translation table 0, then a generation step using generation table
0, followed by two translation steps using translation tables 1 and 2, and finally a generation
step using generation table 1. (The specific names t0, t1, ... are automatically assigned to
translation tables in the order you define them with --translation-factors, and likewise for
g0 etc.)
It is possible to specify multiple decoding paths, for instance by --decoding-steps t0,g0,t1,t2,g1:t3,
where colons separate the paths. Translation options are generated from each decoding path
and used during decoding.
5.4
Training Step 1: Prepare Data
The parallel corpus has to be converted into a format that is suitable to the GIZA++ toolkit. Two
vocabulary files are generated and the parallel corpus is converted into a numberized format.
The vocabulary files contain words, integer word identifiers and word count information:
218
5. Training Manual
==> corpus/de.vcb <==
1
UNK
0
2
,
928579
3
.
723187
4
die
581109
5
der
491791
6
und
337166
7
in
230047
8
zu
176868
9
den
168228
10
ich
162745
==> corpus/en.vcb <==
1
UNK
0
2
the
1085527
3
.
714984
4
,
659491
5
of
488315
6
to
481484
7
and
352900
8
in
330156
9
is
278405
10
that
262619
The sentence-aligned corpus now looks like this:
1
3469 5 2049
4107 5 2 1399
1
10 3214 4 116 2007 2 9 5254 1151 985 6447 2049 21 44 141 14 2580 3
14 2213 1866 2 1399 5 2 29 46 3256 18 1969 4 2363 1239 1111 3
1
7179
306
A sentence pair now consists of three lines: First the frequency of this sentence. In our training
process this is always 1. This number can be used for weighting different parts of the training corpus differently. The two lines below contain word ids of the foreign and the English
sentence. In the sequence 4107 5 2 1399 we can recognize of (5) and the (2).
GIZA++ also requires words to be placed into word classes. This is done automatically by calling the mkcls program. Word classes are only used for the IBM reordering model in GIZA++.
A peek into the foreign word class file:
5.5.
Training Step 2: Run GIZA++
219
!
14
"
14
#
30
%
31
&
10
’
14
(
10
)
14
+
31
,
11
5.5
Training Step 2: Run GIZA++
GIZA++ is a freely available implementation of the IBM models. We need it as a initial step
to establish word alignments. Our word alignments are taken from the intersection of bidirectional runs of GIZA++ plus some additional alignment points from the union of the two
runs.
Running GIZA++ is the most time consuming step in the training process. It also requires a lot
of memory (1-2 GB RAM is common for large parallel corpora).
GIZA++ learns the translation tables of IBM Model 4, but we are only interested in the word
alignment file:
> zcat giza.de-en/de-en.A3.final.gz | head -9
# Sentence pair (1) source length 4 target length 3 alignment score : 0.00643931
wiederaufnahme der sitzungsperiode
NULL ({ }) resumption ({ 1 }) of ({ }) the ({ 2 }) session ({ 3 })
# Sentence pair (2) source length 17 target length 18 alignment score : 1.74092e-26
ich erklaere die am donnerstag , den 28. maerz 1996 unterbrochene sitzungsperiode
des europaeischen parlaments fuer wiederaufgenommen .
NULL ({ 7 }) i ({ 1 }) declare ({ 2 }) resumed ({ }) the ({ 3 }) session ({ 12 })
of ({ 13 }) the ({ }) european ({ 14 }) parliament ({ 15 })
adjourned ({ 11 16 17 }) on ({ }) thursday ({ 4 5 }) , ({ 6 }) 28 ({ 8 })
march ({ 9 }) 1996 ({ 10 }) . ({ 18 })
# Sentence pair (3) source length 1 target length 1 alignment score : 0.012128
begruessung
NULL ({ }) welcome ({ 1 })
220
5. Training Manual
In this file, after some statistical information and the foreign sentence, the English sentence
is listed word by word, with references to aligned foreign words: The first word resumption
({ 1 }) is aligned to the first German word wiederaufnahme. The second word of ({ }) is
unaligned. And so on.
Note that each English word may be aligned to multiple foreign words, but each foreign word
may only be aligned to at most one English word. This one-to-many restriction is reversed in
the inverse GIZA++ training run:
> zcat giza.en-de/en-de.A3.final.gz | head -9
# Sentence pair (1) source length 3 target length 4 alignment score : 0.000985823
resumption of the session
NULL ({ }) wiederaufnahme ({ 1 2 }) der ({ 3 }) sitzungsperiode ({ 4 })
# Sentence pair (2) source length 18 target length 17 alignment score : 6.04498e-19
i declare resumed the session of the european parliament adjourned on thursday ,
28 march 1996 .
NULL ({ }) ich ({ 1 }) erklaere ({ 2 10 }) die ({ 4 }) am ({ 11 })
donnerstag ({ 12 }) , ({ 13 }) den ({ }) 28. ({ 14 }) maerz ({ 15 })
1996 ({ 16 }) unterbrochene ({ 3 }) sitzungsperiode ({ 5 }) des ({ 6 7 })
europaeischen ({ 8 }) parlaments ({ 9 }) fuer ({ }) wiederaufgenommen ({ })
. ({ 17 })
# Sentence pair (3) source length 1 target length 1 alignment score : 0.706027
welcome
NULL ({ }) begruessung ({ 1 })
5.5.1
Training on really large corpora
GIZA++ is not only the slowest part of the training, it is also the most critical in terms of
memory requirements. To better be able to deal with the memory requirements, it is possible
to train a preparation step on parts of the data that involves an additional program called
snt2cooc.
For practical purposes, all you need to know is that the switch --parts n may allow training
on large corpora that would not be feasible otherwise (a typical value for n is 3).
This is currently not a problem for Europarl training, but is necessary for large Arabic and
Chinese training runs.
5.5.2
Training in parallel
Using the --parallel option will fork the script and run the two directions of GIZA++ as
independent processes. This is the best choice on a multi-processor machine.
If you have only single-processor machines and still wish to run the two GIZA++ processes
in parallel, use the following (rather obsolete) trick. Support for this is not fully user friendly,
some manual involvement is essential.
5.6.
Training Step 3: Align Words
221
• First you start training the usual way with the additional switches --last-step 2 --direction
1, which runs the data preparation and one direction of GIZA++ training
• When the GIZA++ step started, start a second training run with the switches --first-step
2 --direction 2. This runs the second GIZA++ run in parallel, and then continues the
rest of the model training. (Beware of race conditions! The second GIZA++ run might
finish earlier than the first one to training step 3 might start too early!)
5.6
Training Step 3: Align Words
To establish word alignments based on the two GIZA++ alignments, a number of heuristics
may be applied. The default heuristic grow-diag-final starts with the intersection of the two
Other possible alignment methods:
•
•
•
•
•
intersection
grow-diag (without final step)
union
srctotgt (only consider word-to-word alignments from the source-target GIZA++ alignment file)
• tgttosrc (only consider word-to-word alignments from the target-source GIZA++ alignment file)
Alternative alignment methods can be specified with the switch --alignment.
Here, the pseudo code for the default heuristic:
GROW-DIAG-FINAL(e2f,f2e):
neighboring = ((-1,0),(0,-1),(1,0),(0,1),(-1,-1),(-1,1),(1,-1),(1,1))
alignment = intersect(e2f,f2e);
GROW-DIAG(); FINAL(e2f); FINAL(f2e);
GROW-DIAG():
iterate until no new points added
for english word e = 0 ... en
for foreign word f = 0 ... fn
if ( e aligned with f )
for each neighboring point ( e-new, f-new ):
if ( ( e-new not aligned or f-new not aligned ) and
( e-new, f-new ) in union( e2f, f2e ) )
add alignment point ( e-new, f-new )
FINAL(a):
for english word e-new = 0 ... en
for foreign word f-new = 0 ... fn
222
5. Training Manual
if ( ( e-new not aligned or f-new not aligned ) and
( e-new, f-new ) in alignment a )
add alignment point ( e-new, f-new )
To illustrate this heuristic, see the example in the Figure below with the intersection of the two
alignments for the second sentence in the corpus above
and then add some additional alignment points that lie in the union of the two alignments
5.6.
Training Step 3: Align Words
223
This alignment has a blatant error: the alignment of the two verbs is mixed up. resumed is
aligned to unterbrochene, and adjourned is aligned to wiederaufgenommen, but it should be
the other way around.
To conclude this section, a quick look into the files generated by the word alignment process:
==> model/aligned.de <==
wiederaufnahme der sitzungsperiode
ich erklaere die am donnerstag , den 28. maerz 1996 unterbrochene sitzungsperiode
des europaeischen parlaments fuer wiederaufgenommen .
begruessung
==> model/aligned.en <==
resumption of the session
i declare resumed the session of the european parliament adjourned on
thursday , 28 march 1996 .
welcome
==> model/aligned.grow-diag-final <==
0-0 0-1 1-2 2-3
0-0 1-1 2-3 3-10 3-11 4-11 5-12 7-13 8-14 9-15 10-2 11-4 12-5 12-6 13-7
14-8 15-9 16-9 17-16
0-0
224
5. Training Manual
The third file contains alignment information, one alignment point at a time, in form of the
position of the foreign and English word.
5.7
Training Step 4: Get Lexical Translation Table
Given this alignment, it is quite straight-forward to estimate a maximum likelihood lexical
translation table. We estimate the w(e|f) as well as the inverse w(f|e) word translation table.
Here are the top translations for europa into English:
> grep ’ europa ’ model/lex.f2n | sort -nrk 3 | head
europe europa 0.8874152
european europa 0.0542998
union europa 0.0047325
it europa 0.0039230
we europa 0.0021795
eu europa 0.0019304
europeans europa 0.0016190
euro-mediterranean europa 0.0011209
europa europa 0.0010586
continent europa 0.0008718
5.8
Training Step 5: Extract Phrases
In the phrase extraction step, all phrases are dumped into one big file. Here is the top of that
file:
wiederaufnahme ||| resumption ||| 0-0
wiederaufnahme der ||| resumption of the ||| 0-0 1-1 1-2
wiederaufnahme der sitzungsperiode ||| resumption of the session ||| 0-0 1-1 1-2 2-3
der ||| of the ||| 0-0 0-1
der sitzungsperiode ||| of the session ||| 0-0 0-1 1-2
sitzungsperiode ||| session ||| 0-0
ich ||| i ||| 0-0
ich erklaere ||| i declare ||| 0-0 1-1
erklaere ||| declare ||| 0-0
sitzungsperiode ||| session ||| 0-0
5.9.
Training Step 6: Score Phrases
225
The content of this file is for each line: foreign phrase, English phrase, and alignment points.
Alignment points are pairs (foreign,english). Also, an inverted alignment file extract.inv
is generated, and if the lexicalized reordering model is trained (default), a reordering file
extract.o.
5.9
Training Step 6: Score Phrases
Subsequently, a translation table is created from the stored phrase translation pairs. The two
steps are separated, because for larger translation models, the phrase translation table does not
fit into memory. Fortunately, we never have to store the phrase translation table into memory
-– we can construct it on disk.
To estimate the phrase translation probability φ(e|f) we proceed as follows: First, the extract file
is sorted. This ensures that all English phrase translations for an foreign phrase are next to each
other in the file. Thus, we can process the file, one foreign phrase at a time, collect counts and
compute φ(e|f) for that foreign phrase f. To estimate φ(f|e), the inverted file is sorted, and then
φ(f|e) is estimated for an English phrase at a time.
Next to phrase translation probability distributions φ(f|e) and φ(e|f), additional phrase translation scoring functions can be computed, e.g. lexical weighting, word penalty, phrase penalty,
etc. Currently, lexical weighting is added for both directions and a fifth score is the phrase
penalty.
> grep ’| in europe |’ model/phrase-table | sort -nrk 7 -t\| | head
in europa ||| in europe ||| 0.829007 0.207955 0.801493 0.492402
europas ||| in europe ||| 0.0251019 0.066211 0.0342506 0.0079563
in der europaeischen union ||| in europe ||| 0.018451 0.00100126 0.0319584 0.0196869
in europa , ||| in europe ||| 0.011371 0.207955 0.207843 0.492402
europaeischen ||| in europe ||| 0.00686548 0.0754338 0.000863791 0.046128
im europaeischen ||| in europe ||| 0.00579275 0.00914601 0.0241287 0.0162482
fuer europa ||| in europe ||| 0.00493456 0.0132369 0.0372168 0.0511473
in europa zu ||| in europe ||| 0.00429092 0.207955 0.714286 0.492402
an europa ||| in europe ||| 0.00386183 0.0114416 0.352941 0.118441
der europaeischen ||| in europe ||| 0.00343274 0.00141532 0.00099583 0.000512159
Currently, four different phrase translation scores are computed:
1.
2.
3.
4.
inverse phrase translation probability φ(f|e)
inverse lexical weighting lex(f|e)
direct phrase translation probability φ(e|f)
direct lexical weighting lex(e|f)
Previously, there was another score:
1. phrase penalty (always exp(1) = 2.718)
This has now been superceded by it’s own feature function, PhrasePenalty.
226
5. Training Manual
Using a subset of scores
You may not want to use all the scores in your translation table. The following options allow
you to remove some of the scores:
• NoLex -- do not use lexical scores (removes score 2 and 4)
• OnlyDirect -- do not use the inverse scores (removes score 1 and 2)
These settings have to be specified with the setting -score-options when calling the script
train-model.perl, for instance:
train-model.perl [... other settings ...] -score-options ’--NoLex’
NB - the consolidate program (that runs after score) also has a few arguments. For example, it
has
• PhraseCount -- add the old phrase count feature (score 5)
However, this can’t be set with by the train-model.perl script.
Good Turing discounting
Singleton phrase pairs tend to have overestimated phrase translation probabilities. Consider
the extreme case of a source phrase that occurs only once in the corpus and has only one translation. The corresponding phrase translation probability φ(e|f) would be 1.
To obtain better phrase translation probabilities, the observed counts may be reduced by expected counts which takes unobserved events into account. Borrowing a method from language
model estimation, Good Turing discounting can be used to reduce the actual counts (such as
1 in the example above) and reduce it to a more realistic number (maybe 0.3). The value of
the adjusted count is determined by an analysis of the number of singleton, twice-occuring,
thrice-occuring, etc. phrase pairs that were extracted.
To use Good Turing discounting of the phrase translation probabilities, you have to specify
--GoodTuring as one of the -score-options, as in the section above. The adjusted counts are
reported to STDERR.
Word-to-word alignment
An enhanced version of the scoring script outputs the word-to-word alignments between f and
e as they are in the files (extract and extract.inv) generated in the previous training step
"Extract Phrases" (Section 5.8).
The alignments information are reported in the fourth fields. The format is identical to the
alignment output obtained when the GIZA++ output has been symmetrized priot to phrase
extraction.
5.9.
Training Step 6: Score Phrases
227
> grep ’| in europe |’ model/phrase-table | sort -nrk 7 -t\| | head
in europa ||| in europe ||| 0.829007 0.207955 ||| 0-0 1-1 ||| ...
europas ||| in europe ||| ... ||| 0-0 0-1 ||| ...
in der europaeischen union ||| in europe ||| ... ||| 0-0 2-1 3-1 |||
in europa , ||| in europe ||| ... ||| 0-0 1-1 ||| ...
europaeischen ||| in europe ||| ... ||| 0-1 ||| ...
im europaeischen ||| in europe ||| ... ||| 0-0 1-1 |||
For instance:
in der europaeischen union ||| in europe ||| 0-0 2-1 3-1 ||| ...
means
German
in
der
europaeischen
union
->
->
->
->
->
English
in
europe
europe
The word-to-word alignments come from one word alignment (see training step "Align words"
(Section 5.6)).
The alignment information is also used in SCFG-rules for the chart-decoder to link non-terminals
together in the source and target side. In this instance, the alignment information is not an option, but a necessity. For example, the following Moses SCFG rule
[X][X] miss [X][X] [X] ||| [X][X] [X][X] manques [X] ||| ... ||| 0-1 2-0 ||| ...
is formated as this in the Hiero format:
[X] ||| [X,1] miss [X,2] ||| [X,2] [X,1] manques ||| ....
ie. this rule reordes the 1st and 3rd non-terminals in the source.
Therefore, the same alignment field can be used for word-alignment and non-terminal coindexes. However, I’m (Hieu) sure if anyone has implemented this in the chart decoder yet
228
5. Training Manual
5.10
Training Step 7: Build reordering model
By default, only a distance-based reordering model is included in final configuration. This
model gives a cost linear to the reordering distance. For instance, skipping over two words
costs twice as much as skipping over one word.
However, additional conditional reordering models, so called lexicalized reordering models,
may be build. There are three types of lexicalized reordering models in Moses that are based
on Koehn et al. (2005)1 and Galley and Manning (2008)2 . The Koehn at al. model determines
the orientation of two phrases based on word alignments at training time, and based on phrase
alignments at decoding time. The other two models are based on Galley and Manning. The
phrase-based model uses phrases both at training and decoding time, and the hierarchical
model allows combinations of several phrases for determining the orientation.
The lexicalized reordering models are specified by a configuration string, containing five parts,
that account for different aspects:
• Modeltype - the type of model used (see above)
– wbe - word-based extraction (but phrase-based at decoding). This is the original
model in Moses. DEFAULT
– phrase - phrase-based model
– hier - hierarchical model
• Orientation - Which classes of orientations that are used in the model
– mslr - Considers four different orientations: monotone, swap, discontinuous-left,
discontinuous-right
– msd - Considers three different orientations: monotone, swap, discontinuous (the
two discontinuous classes of the mslr model are merged into one class)
– monotonicity - Considers two different orientations: monotone or non-monotone
(swap and discontinuous of the msd model are merged into the non-monotone class)
– leftright - Considers two different orientations: left or right (the four classes
in the mslr model are merged into two classes, swap and discontinuous-left into
left and monotone and discontinuous-right into right)
• Directionality - Determines if the orientation should be modeled based on the previous
or next phrase, or both.
– backward - determine orientation with respect to previous phrase DEFAULT
– forward - determine orientation with respect to following phrase
– bidirectional - use both backward and forward models
• language - decides which language to base the model on
– fe - conditioned on both the source and target languages
– f - conditioned on the source language only
• collapsing - determines how to treat the scores
– allff - treat the scores as individual feature functions DEFAULT
– collapseff - collapse all scores in one direction into one feature function
any possible configuration of these five factors is allowed. It is always necessary to specify
orientation and language. The other three factors use the default values indicated above if they
are not specified. Some examples of possible models are:
1
2
http://homepages.inf.ed.ac.uk/pkoehn/publications/iwslt05-report.pdf
http://www.aclweb.org/anthology/D/D08/D08-1089.pdf
5.11.
Training Step 8: Build generation model
229
• msd-bidirectional-fe (this model is commonly used, for instance it is the model used
in the WMT baselines3 )
• wbe-msd-bidirectional-fe-allff same model as above
• mslr-f
• wbe-backward-mslr-f-allff same model as above
• phrase-msd-bidirectional-fe
• hier-mslr-bidirectional-fe
• hier-leftright-forward-f-collapseff
and of course distance.
Which reordering model(s) that are used (and built during the training process, if necessary)
can be set with the switch -reordering, e.g.:
-reordering
-reordering
-reordering
-reordering
distance
msd-bidirectional-fe
msd-bidirectional-fe,hier-mslr-bidirectional-fe
distance,msd-bidirectional-fe,hier-mslr-bidirectional-fe
Note that the distance model is always included, so there is no need to specify it.
The number of features that are created with a lexical reordering model depends on the type of
the model. If the flag allff is used, a msd model has three features, one each for the probability
that the phrase is translated monotone, swapped, or discontinuous, a mslr model has four
features and a monotonicity or leftright model has two features. If a bidirectional model
is used, then the number of features doubles - one for each direction. If collapseff are used
there is one feature for each direction, regardless of which orientation types that are used.
There are also a number of other flags that can be given to train-model.perl that concerns the
reordering models:
• --reordering-smooth - specifies the smoothing constant to be used for training lexicalized reordering models. If the letter u follows the constant, smoothing is based on actual
counts. (default 0.5)
• --max-lexical-reordering - if this flag is used, the extract file will contain information
for the mslr orientations for all three model types, wbe, phrase and hier. Otherwise
the extract file will contain the minimum information that is needed based on which
reordering model config strings that are given.
5.11
Training Step 8: Build generation model
The generation model is build from the target side of the parallel corpus.
By default, forward and backward probabilities are computed. If you use the switch --generation-type
single only the probabilities in the direction of the step are computed.
3
http://www.statmt.org/wmt11/baseline.html
230
5. Training Manual
5.12
Training Step 9: Create Configuration File
As a final step, a configuration file for the decoder is generated with all the correct paths for
the generated model and a number of default parameter settings.
This file is called model/moses.ini
You will also need to train a language model. This is described in the decoder manual.
Note that the configuration file set --by default-- the usage of SRILM as a LM toolkit. If you
prefer to use another LM toolkit, change the configuration file as described here4
5.13
Building a Language Model
5.13.1
Language Models in Moses
The language model should be trained on a corpus that is suitable to the domain. If the translation model is trained on a parallel corpus, then the language model should be trained on the
output side of that corpus, although using additional training data is often beneficial.
Our decoder works with the following language models:
•
•
•
•
•
the SRI language modeling toolkit5 , which is freely available.
the IRST language modeling toolkit6 , which is freely available and open source.
the RandLM language modeling toolkit7 , which is freely available and open source.
the KenLM language modeling toolkit8 , which is included in Moses by default.
the DALM language modeling toolkit9 , which is freely available and open source.
To use these language models, they have to be compiled with the proper option:
•
•
•
•
--with-srilm=<root dir of the SRILM toolkit>
--with-irstlm=<root dir of the IRSTLM toolkit>
--with-randlm=<root dir of the RandLM toolkit>
--with-dalm=<root dir of the DALM toolkit>
KenLM is compiled by default. In the Moses configuration file, the type (SRI/IRST/RandLM/KenLM/DALM)
of the LM is specified by the feature function class, eg.
[feature]
SRILM path=filename.srilm order=3 .....
4
http://www.statmt.org/moses/?n=FactoredTraining.BuildingLanguageModel#ntoc1
http://www.speech.sri.com/projects/srilm/
6
http://sourceforge.net/projects/irstlm/
7
http://sourceforge.net/projects/randlm/
8
http://kheafield.com/code/kenlm/
9
https://github.com/jnory/DALM
5
5.13.
Building a Language Model
231
or
IRSTLM path=filename.irstlm ...
or
RANDLM path=filename.irstlm ...
or
KENLM path=filename.arpa ...
or
DALM path=filename.dalm ...
The toolkits all come with programs that create a language model file, as required by our decoder. ARPA files are generally exchangeable, so you can estimate with one toolkit and query
with a different one.
5.13.2
Building a LM with the SRILM Toolkit
A language model can be created by calling:
ngram-count -text CORPUS_FILE -lm SRILM_FILE
The command works also on compressed (gz) input and output. There are a variety of switches
that can be used, we recommend -interpolate -kndiscount.
232
5. Training Manual
5.13.3
On the IRSTLM Toolkit
Moses can also use language models created with the IRSTLM toolkit (see Federico & Cettolo, (ACL WS-SMT, 2007)10 ). The commands described in the following are supplied with the
IRSTLM toolkit handles LM formats which permit to reduce both storage and decoding memory requirements, and to save time in LM loading. In particular, it provides tools for:
•
•
•
•
•
building (huge) LMs (Section 5.13.3)
quantizing LMs (Section 5.13.3)
compiling LMs (possibly quantized) into a binary format (Section 5.13.3)
accessing binary LMs through the memory mapping mechanism (Section 5.13.3)
query class and chunk LMs (Section 5.13.3)
Compiling IRSTLM
Compiling IRSTLM requires:
1. automake 1.9 or higher
2. autoconf 2.59 or higher
3. libtool 2.2.6 or higher
This is the commands I (Hieu) executed to download and install IRSTLM:
tar zxvf irstlm-5.80.03.tgz
cd irstlm-5.80.03/
./regenerate-makefiles.sh
./configure --prefix=$PWD make -j8 make install This the binaries and libraries in the current directory, in bin/ and lib/. Note - as of January 2014, there are problems with the source code in the IRSTLM sourceforge repository. Please use the version 5.80.03 available for download on the website12 . 10 http://www.aclweb.org/anthology-new/W/W07/W07-0712.pdf http://sourceforge.net/projects/irstlm 12 http://sourceforge.net/projects/irstlm/ 11 5.13. Building a Language Model 233 Building Huge Language Models Training a language model from huge amounts of data can be definitively memory and time expensive. The IRSTLM toolkit features algorithms and data structures suitable to estimate, store, and access very large LMs. IRSTLM is open source and can be downloaded from here13 . Typically, LM estimation starts with the collection of n-grams and their frequency counters. Then, smoothing parameters are estimated for each n-gram level; infrequent n-grams are possibly pruned and, finally, a LM file is created containing n-grams with probabilities and back-off weights. This procedure can be very demanding in terms of memory and time if applied to huge corpora. IRSTLM provides a simple way to split LM training into smaller and independent steps, which can be distributed among independent processes. The procedure relies on a training script that makes little use of computer memory and implements the Witten-Bell smoothing method. (An approximation of the modified Kneser-Ney smoothing method is also available.) First, create a special directory stat under your working directory, where the script will save lots of temporary files; then, simply run the script buildlm.sh as in the example: build-lm.sh -i "gunzip -c corpus.gz" -n 3 -o train.irstlm.gz -k 10 The script builds a 3-gram LM (option -n) from the specified input command (-i), by splitting the training procedure into 10 steps (-k). The LM will be saved in the output (-o) file train.irstlm.gz with an intermediate ARPA format. This format can be properly managed through the compile-lm command in order to produce a compiled version or a standard ARPA version of the LM. For a detailed description of the procedure and of other commands available under IRSTLM please refer to the user manual supplied with the package. Binary Language Models You can convert your language model file (created either with the SRILM ngram-count command or with the IRSTLM toolkit) into a compact binary format with the command: compile-lm language-model.srilm language-model.blm Moses compiled with the IRSTLM toolkit is able to properly handle that binary format; the setting of moses.ini for that file is: 13 http://sourceforge.net/projects/irstlm 234 5. Training Manual IRSTLM order=3 factor=0 path=language-model.blm The binary format allows LMs to be efficiently stored and loaded. The implementation privileges memory saving rather than access time. Quantized Language Models Before compiling the language model, you can quantize (see Federico & Bertoldi, (ACL WSSMT, 2006)14 ) its probabilities and back-off weights with the command: quantize-lm language-model.srilm language-model.qsrilm Hence, the binary format for this file is generated by the commmand: compile-lm language-model.qsrilm language-model.qblm The resulting language model requires less memory because all its probabilities and back-off weights are now stored in 1 byte instead of 4. No special setting of the configuration file is required: Moses compiled with the IRSTLM toolkit is able to read the necessary information from the header of the file. Memory Mapping It is possible to avoid the loading of the LM into the central memory by exploiting the memory mapping mechanism. Memory mapping permits the decoding process to directly access the (binary) LM file stored on the hard disk. Warning: In case of parallel decoding in a cluster of computers, each process will access the same file. The possible large number of reading requests could overload the driver of the hard disk which the LM is stored on, and/or the network. One possible solution to such a problem is to store a copy of the LM on the local disk of each processing node, for example under the /tmp/ directory. In order to activate the access through the memory mapping, simply add the suffix .mm to the name of the LM file (which must be stored in the binary format) and update the Moses configuration file accordingly. As an example, let us suppose that the 3gram LM has been built and stored in binary format in the file 14 http://www.aclweb.org/anthology/W/W06/W06-3113 5.13. Building a Language Model 235 language-model.blm Rename it for adding the .mm suffix: mv language-model.blm language-model.blm.mm or create a properly named symbolic link to the original file: ln -s language-model.blm language-model.blm.mm Now, the activation of the memory mapping mechanism is obtained simply by updating the Moses configuration file as follows: IRSTLM order=3 factor=0 path=language-model.blm.mm Class Language Models and more Typically, LMs employed by Moses provide the probability of n-grams of single factors. In addition to the standard way, the IRSTLM toolkit allows Moses to query the LMs in other different ways. In the following description, it is assumed that the target side of training texts contains words which are concatenation of N>=1 fields separated by the character #. Similarly to factored models, where the word is not anymore a simple token but a vector of factors that can represent different levels of annotation, here the word can be the concatenation of different tags for the surface form of a word, e.g.: word#lemma#part-of-speech#word-class Specific LMs for each tag can be queried by Moses simply by adding a fourth parameter in the line of the configuration file devoted to the specification of the LM. The additional parameter is a file containing (at least) the following header: 236 5. Training Manual FIELD <int> Possibly, it can also include a one-to-one map which is applied to each component of n-grams before the LM query: w1 class(w1) w2 class(w2) ... wM class(wM) The value of <int> determines the processing applied to the n-gram components, which are supposed to be strings like field0#field1#...#fieldN: • -1: the strings are used are they are; if the map is given, it is applied to the whole string before the LM query • 0-9: the field number <int> is selected; if the map is given, it is applied to the selected field • 00-99: the two fields corresponding to the two digits are selected and concatenated together using the character _ as separator. For example, if <int>=21, the LM is queried with n-grams of strings field2_field1. If the map is given, it is applied to the field corresponding to the first digit. The last case is useful for lexicalization of LMs: if the fields n. 2 and 1 correspond to the POS and lemma of the actual word respectively, the LM is queried with n-grams of POS_lemma. Chunk Language Models A particular processing is performed whenever fields are supposed to correspond to microtags, i.e. the per-word projections of chunk labels. The processing aims at collapsing the sequence of microtags defining a chunk to the label of that chunk. The chunk LM is then queried with n-grams of chunk labels, in an asynchronous manner with respect to the sequence of words, as in general chunks consist of more words. The collapsing operation is automatically activated if the sequence of microtags is: (TAG TAG+ TAG+ ... TAG+ TAG) or TAG( TAG+ TAG+ ... TAG+ TAG) 5.13. Building a Language Model 237 Both those sequences are collapsed into a single chunk label (let us say CHNK) as long as (TAG / TAG(, TAG+ and TAG) are all mapped into the same label CHNK. The map into different labels or a different use/position of characters (, + and ) in the lexicon of tags prevent the collapsing operation. Currently (Aug 2008), lexicalized chunk LMs are still under investigation and only non-lexicalized chunk LMs are properly handled; then, the range of admitted <int> values for this kind of LMs is -1...9, with the above described meaning. 5.13.4 RandLM If you really want to build the largest LMs possible (for example, a 5-gram trained on one hundred billion words then you should look at the RandLM. This takes a very different approach to either the SRILM or the IRSTLM. It represents LMs using a randomized data structure (technically, variants of Bloom filters). This can result in LMs that are ten times smaller than those created using the SRILM (and also smaller than IRSTLM), but at the cost of making decoding about four times slower. RandLM is multithreaded now, so the speed reduction should be less of a problem. Technical details of randomized language modelling can be found in a ACL paper (see Talbot and Osborne, (ACL 2007)15 ) Installing RandLM RandLM is available at Sourceforge16 . After extracting the tar ball, go to the directory src and type make. For integrating RandLM into Moses, please see above. Building a randomized language model The buildlm binary (in randlm/bin) preprocesses and builds randomized language models. The toolkit provides three ways for building a randomized language models: 1. from a tokenised corpus (this is useful for files around 100 million words or less) 2. from a precomputed backoff language model in ARPA format (this is useful if you want to use a precomputed SRILM model) 3. from a set of precomputed ngram-count pairs (this is useful if you need to build LMs from billions of words. RandLM has supporting Hadoop scripts). The former type of model will be referred to as a CountRandLM while the second will be referred to as a BackoffRandLM. Models built from precomputed ngram-count pairs are also of type "CountRandLM". CountRandLMs use either StupidBackoff or else Witten-Bell smoothing. BackoffRandLM models can use any smoothing scheme that the SRILM implements. Generally, CountRandLMs are smaller than BackoffRandLMs, but use less sophisticated smoothing. 15 16 http://aclweb.org/anthology-new/P/P07/P07-1065.pdf http://sourceforge.net/projects/randlm/ 238 5. Training Manual When using billions of words of training material there is less of a need for good smoothing and so CountRandLMs become appropriate. The following parameters are important in all cases: • struct: The randomized data structure used to represent the language model (currently only BloomMap and LogFreqBloomFilter). • order: The order of the n-gram model e.g., 3 for a trigram model. • falsepos: The false positive rate of the randomized data structure on an inverse log scale so -falsepos 8 produces a false positive rate of 1/28 . • values: The quantization range used by the model. For a CountRandLM quantisation is performed by taking a logarithm. The base of the logarithm is set as 2<sup>1/values</sup>. For a BackoffRandLM a binning quantisation algorithm is used. The size of the codebook is set as 2<sup>values</sup>. A reasonable setting in both cases is -values 8. • input-path: The location of data to be used to create the language model. • input-type: The format of the input data. The following four formats are supported – for a CountRandLM: ∗ corpus tokenised corpora one sentence per line; ∗ counts n-gram counts file (one count and one n-gram per line); – Given a ’corpus’ file the toolkit will create a ’counts’ file which may be reused (see examples below). – for a BackoffRandLM: ∗ arpa an ARPA backoff language model; ∗ backoff language model file (two floats and one n-gram per line). – Given an arpa file the toolkit will create a ’backoff’ file which may be reused (see examples below). • output-prefix:Prefix added to all output files during the construction of a randomized language model. Example 1: Building directly from corpora The command ./buildlm -struct BloomMap -falsepos 8 -values 8 -output-prefix model -order 3 < corpus would produce the following files:- model.BloomMap model.counts.sorted model.stats model.vcb <<<<- the randomized language model n-gram counts file statistics file (counts of counts) vocabulary file (not needed) model.BloomMap: This randomized language model is ready to use on its own (see ’Querying a randomized language model’ below). 5.13. Building a Language Model 239 model.counts.sorted: This is a file in the RandLM ’counts’ format with one count followed by one n-gram per line. It can be specified as shown in Example 3 below to avoid recomputation when building multiple randomized language models from the same corpus. model.stats: This statistics file contains counts of counts and can be specified via the optional parameter ’-statspath’ as shown in Example 3 to avoid recomputation when building multiple randomized language models from the same data. Example 2: Building from an ARPA file (from another toolkit) The command ./buildlm -struct BloomMap -falsepos 8 -values 8 -output-prefix model -order 3 \ -input-path precomputed.bo -input-type arpa (where precomputed.bo contains an ARPA-formatted backoff model) would produce the following files: model.BloomMap <- the randomized language model model.backoff <- RandLM backoff file model.stats <- statistics file (counts of counts) model.vcb <- vocabulary file (not needed) model.backoff is a RandLM formatted copy of the ARPA model. It can be reused in the same manner as the model.counts.sorted file (see Example 3). Example 3: Building a second randomized language model from the same data The command ./buildlm -struct BloomMap -falsepos 4 -values 8 -output-prefix model4 -order 3 -input-path model.counts.sorted -input-type counts -stats-path model.stats would construct a new randomized language model (model4.BloomMap) from the same data as used in Example 1 but with a different error rate (here -falsepos 4). This usage avoids re-tokenizing the corpus and recomputing the statistics file. 240 5. Training Manual Building Randomised LMs from 100 Billion Words using Hadoop At some point you will discover that you cannot build a LM using your data. RandLM natively uses a disk-based method for creating n-grams and counts, but this will be slow for large corpora. Instead you can create these ngram-count pairs using Hadoop (Map-Reduce). The RandLM release has Hadoop scripts which take raw text files and create ngram-counts. We have built randomised LMs this way using more than 110 billion tokens. The procedure for using Hadoop is as follows: • You first load raw and possibly tokenised text files onto the Hadoop Distributed File System (DFS). This will probably involve commands such as: Hadoop dfs -put myFile data/ • You then create ngram-counts using Hadoop (here a 5-gram): perl hadoop-lm-count.prl data data-counts 5 data-counting • You then upload the counts to the Unix filesystem: perl hadoopRead.prl data-counts | gzip - > /unix/path/to/counts.gz • These counts can then be passed to RandLM: ./buildlm -estimator batch -smoothing WittenBell -order 5 \ -values 12 -struct LogFreqBloomFilter -tmp-dir /disk5/miles \ -output-prefix giga3.rlm -output-dir /disk5/miles -falsepos 12 \ -keep-tmp-files -sorted-by-ngram -input-type counts \ -input-path /disk5/miles/counts.gz Querying Randomised Language Models Moses uses its own interface to the randLM, but it may be interesting to query the language model directly. The querylm binary (in randlm/bin) allows a randomized language model to be queried. Unless specified the scores provided by the tool will be conditional log probabilities (subject to randomisation errors). The following parameters are available:- 5.13. Building a Language Model 241 • randlm: The path of the randomized language model built using the buildlm tool as described above. • test-path: The location of test data to be scored by the model. • test-type: The format of the test data: currently corpus and ngrams are supported. corpus will treat each line in the test file as a sentence and provide scores for all n-grams (adding <s> and </s>). ngrams will score each line once treating each as an independent n-gram. • get-counts: Return the counts of n-grams rather than conditional log probabilities (only supported by CountRandLM). • checks: Applies sequential checks to n-grams to avoid unnecessary false positives. Example: The command ./querylm -randlm model.BloomMap -test-path testfile -test-type ngrams -order 3 > scores would write out conditional log probabilities one for each line in the file test-file. • Finally, you then tell randLM to use these pre-computed counts. 5.13.5 KenLM KenLM is a language model that is simultaneously fast and low memory. The probabilities returned are the same as SRI, up to floating point rounding. It is maintained by Ken Heafield, who provides additional information on his website17 , such as benchmarks18 comparing speed and memory use against the other language model implementations. KenLM is distributed with Moses and compiled by default. KenLM is fully thread-safe for use with multi-threaded Moses. Estimation The lmplz program estimates language models with Modified Kneser-Ney smoothing and no pruning. Pass the order (-o), an amount of memory to use for building (-S), and a location to place temporary files (-T). Note that -S is compatible with GNU sort so e.g. 1G = 1 gigabyte and 80% means 80% of physical RAM. It scales to much larger models than SRILM can handle and does not resort to approximation like IRSTLM does. bin/lmplz -o 5 -S 80% -T /tmp <text >text.arpa See the page on estimation19 for more. 17 http://kheafield.com/code/kenlm/ http://kheafield.com/code/kenlm/benchmark/ 19 http://kheafield.com/code/kenlm/estimation/ 18 242 5. Training Manual Using the EMS To use lmplz in EMS set the following three parameters to your needs and copy the fourth one as is. # path to lmplz binary lmplz =$moses-bin-dir/lmplz
# order of the language model
order = 3
# additional parameters to lmplz (check lmplz help message)
settings = "-T $working-dir/tmp -S 10G" # this tells EMS to use lmplz and tells EMS where lmplz is located lm-training = "$moses-script-dir/generic/trainlm-lmplz.perl -lmplz $lmplz" Querying ARPA files can be read directly: KENLM factor=<factor> order=<order> path=filename.arpa but the binary format loads much faster and provides more flexibility. The <order> field is ignored. By contrast, SRI silently returns incorrect probabilities if you get it wrong (KneserNey smoothed probabilties for lower-order n-grams are conditioned on backing off). Binary file Using the binary format significantly reduces loading time. It also exposes more configuration options. The kenlm/build_binary program converts ARPA files to binary files: kenlm/build_binary filename.arpa filename.binary This will build a binary file that can be used in place of the ARPA file. Note that, unlike IRST, the file extension does not matter; the binary format is recognized using magic bytes. You can also specify the data structure to use: kenlm/build_binary trie filename.arpa filename.binary where valid values are probing, sorted, and trie. The default is probing. Generally, I recommend using probing if you have the memory and trie if you do not. See benchmarks for details. To determine the amount of RAM each data structure will take, provide only the arpa file: 5.13. Building a Language Model 243 kenlm/build_binary filename.arpa Bear in mind that this includes only language model size, not the phrase table or decoder state. Building the trie entails an on-disk sort. You can optimize this by setting the sorting memory with -S using the same options as GNU sort e.g. 100M, 1G, 80%. Final model building will still use the amount of memory needed to store the model. The -T option lets you customize where to place temporary files (the default is based on the output file name). kenlm/build_binary -T /tmp/trie -S 1G trie filename.arpa filename.binary Full or lazy loading KenLM supports lazy loading via mmap. This allows you to further reduce memory usage, especially with trie which has good memory locality. This is specified by another arguments in the feature function for the KENLM feature function: KENLM ... lazyken=<true/false> I recommend fully loading if you have the RAM for it; it actually takes less time to load the full model and use it because the disk does not have to seek during decoding. Lazy loading works best with local disk and is not recommended for networked filesystems. Probing Probing is the fastest and default data structure. Unigram lookups happen by array index. Bigrams and longer n-grams are hashed to 64-bit integers which have very low probability of collision, even with the birthday attack20 . This 64-bit hash is the key to a probing hash table where values are probability and backoff. A linear probing hash table is an array consisting of blanks (zeros) and entries with non-zero keys. Lookup proceeds by hashing the key modulo the array size, starting at this point in the array, and scanning forward until the entry or a blank is found. The ratio of array size to number of entries is controlled by the probing multiplier parameter p. This is a time-space tradeoff: space is linear in p and time is O(p/(p-1)). The value of p can be set at binary building time e.g. kenlm/build_binary -p 1.2 probing filename.arpa filename.binary sets a value of 1.2. The default value is 1.5 meaning that one third of the array is blanks. 20 http://en.wikipedia.org/wiki/Birthday_attack 244 5. Training Manual Trie The trie data structure uses less memory than all other options (except RandLM with stupid backoff), has the best memory locality, and is still faster than any other toolkit. However, it does take longer to build. It works in much the same way as SRI and IRST’s inverted option. Like probing, unigram lookup is an array index. Records in the trie have a word index, probability, backoff, and pointer. All of the records for n-grams of the same order are stored consecutively in memory. An n-gram’s pointer is actually the index into the (n+1)-gram array where block of (n+1)-grams with one more word of history starts. The end of this block is found by reading the next entry’s pointer. Records within the block are sorted by word index. Because the vocabulary ids are randomly permuted, a uniform key distribution applies. Interpolation search within each block finds the word index and its correspoding probability, backoff, and pointer. The trie is compacted by using the minimum number of bits to store each integer. Probability is always non-positive, so the sign bit is also removed. Since the trie stores many vocabulary ids and uses the minimum number of bits to do so, vocabulary filtering is highly effective for reducing overall model size even if less n-grams of higher order are removed. Quantization The trie supports quantization to any number of bits from 1 to 25. To quantize to 8 bits, use -q 8. If you want to separately control probability and backoff quantization, use -q for probability and -b for backoff. Array compression (also known as Chop) The trie pointers comprise a sorted array. These can be compressed using a technique from Raj and Whittaker by chopping off bits and storing offsets instead. The -a option acts as an upper bound on the number of bits to chop; it will never chop more bits than minimizes memory use. Since this is a time-space tradeoff (time is linear in the number of bits chopped), you can set the upper bound number of bits to chop using -a. To minimize memory, use -a 64. To save time, specify a lower limit e.g. -a 10. Vocabulary lookup The original strings are kept at the end of the binary file and passed to Moses at load time to obtain or generate Moses IDs. This is why lazy binary loading still takes a few seconds. KenLM stores a vector mapping from Moses ID to KenLM ID. The cost of this vector and Moses-side vocabulary word storage are not included in the memory use reported by build_binary. However, benchmarks21 report the entire cost of running Moses. Subsection last modified on January 13, 2014, at 01:46 PM 5.14 Tuning 5.14.1 Overview During decoding, Moses scores translation hypotheses using a linear model. In the traditional approach, the features of the model are the probabilities from the language models, phrase/rule tables, and reordering models, plus word, phrase and rule counts. Recent versions of Moses 21 http://kheafield.com/code/kenlm/benchmark/ 5.14. Tuning 245 support the augmentation of these core features with sparse features (Section 7.7), which may be much more numerous. Tuning refers to the process of finding the optimal weights for this linear model, where optimal weights are those which maximise translation performance on a small set of parallel sentences (the tuning set). Translation performance is usually measured with Bleu, but the tuning algorithms all support (at least in principle) the use of other performance measures. Currently (July 2013) only the MERT implementation supports any metrics other than Bleu - it has support for TER, PER CDER and others as well as support for interpolations of metrics. The interest in sparse features has led to the development of new tuning algorithms, and Moses contains implementations of some of these. There are essentially two classes of tuning algorithms used in statistical MT: batch and online. Examples of each of these classes of algorithms are listed in the following sections. 5.14.2 Batch tuning algorithms Here the whole tuning set is decoded, usually generating an n-best list or a lattice, then the model weights are updated based on this decoder output. The tuning set is then re-decoded with the new weights, the optimisation repeated, and this iterative process continues until some convergence criterion is satisfied. All the batch algorithms in Moses are controlled by the inaccurately named mert-moses.pl, which runs the ’outer loop’ (i.e. the repeated decodes). Running this script with no arguments displays usage information. MERT Minimum error rate training (MERT) was introduced by Och (2003)22 . For details on the Moses implementation, see Bertoldi et al, (2009)23 . This line-search based method is probably still the most widely used tuning algorithm, and the default option in Moses. It does not support the use of more than about 20-30 features, so for sparse features you should use one of the other algorithms. Lattice MERT A variant of MERT which uses lattices instead of n-best lists. This was implemented by Kārlis Goba and Christian Buck at the Fourth Machine Translation Marathon in January 2010. It is based on the work of Macherey et al. (2008)24 and is available here25 . PRO Pairwise ranked optimization (Hopkins and May, 2011)26 works by learning a weight set that ranks translation hypotheses in the same order as the metric (e.g. Bleu). Passing the argument --pairwise-ranked to mert-moses.pl enables PRO. 22 http://aclweb.org/anthology-new/P/P03/P03-1021.pdf http://homepages.inf.ed.ac.uk/bhaddow/prague-mert.pdf 24 http://research.google.com/pubs/pub34629.html 25 https://github.com/christianbuck/Moses-Lattice-MERT 26 http://www.aclweb.org/anthology/D11-1125.pdf 23 246 5. Training Manual Batch MIRA Also known as k-best MIRA (Cherry and Foster, 2012)27 , this is a version of MIRA (a marginbased classification algorithm) which works within a batch tuning framework. To use batch MIRA, you need to pass the --batch-mira argument to mert-moses.pl. See below (Section 5.14.4) for more detail. 5.14.3 Online tuning algorithms These methods requires much tighter integration with the decoder. Each sentence in the tuning set is decoded in turn, and based on the results of the decode the weights are updated before decoding the next sentence. The algorithm may iterate through the tuning set multiple times. MIRA The MIRA tuning algorithm (Chiang, 2012)28 ; (Hasler et al, 2011)29 was inspired by the passiveaggressive algorithms of Koby Crammer, and their application to structured prediction by Ryan MacDonald. After decoding each sentence, MIRA updates the weights only if the metric ranks the output sentence with respect to a (pseudo-)reference translation differently from the model. 5.14.4 Tuning in Practice Multiple references To specify multiple references to mert-moses.pl, name each reference file with a prefix followed by a number. Pass the prefix as the reference and ensure that the prefix does not exist. ZMERT Tuning Kamil Kos created contrib/zmert-moses.pl, a Java replacement for mert-moses.pl for those who wish to use ZMERT. The zmert-moses.pl script supports most of the mert-moses.pl parameters, therefore the transition to the new zmert version should be relatively easy. For more details on supported parameters run zmert-moses.pl --help. ZMERT can support multiple metrics ZMERT homepage30 . For instance, SemPOS31 which is based on the tectogrammatical layer, see TectoMT32 . ZMERT JAR, version 1.41 needs to be downloaded from Omar Zaidan’s website33 . If you would like to add a new metric, please, modify the zmert/zmert.jar file in the following way: 27 http://aclweb.org/anthology-new/N/N12/N12-1047.pdf http://www.jmlr.org/papers/volume13/chiang12a/chiang12a.pdf 29 http://ufal.mff.cuni.cz/pbml/96/art-hasler-haddow-koehn.pdf 30 http://www.cs.jhu.edu/~ozaidan/zmert/ 31 http://ufal.mff.cuni.cz/pbml/92/art-pbml92-kos-bojar.pdf 32 http://ufal.mff.cuni.cz/tectomt 33 http://www.cs.jhu.edu/~ozaidan/zmert/ 28 5.14. 1. 2. 3. 4. Tuning 247 extract zmert.jar content by typing jar xf zmert.jar modify the files (probably a copy of NewMetric.java.template) recompile java files by javac *.java create a new version of zmert.jar by typing jar cvfM zmert.jar *.java* *.class k-best batch MIRA Tuning This is hope-fear MIRA built as a drop-in replacement for MERT; it conducts online training using aggregated k-best lists as an approximation to the decoder’s true search space. This allows it to handle large features, and it often out-performs MERT once feature counts get above 10. You can tune using this system by adding --batch-mira to your mert-moses.pl command. This replaces the normal call to the mert executable with a call to kbmira. I recommend also adding the flag --return-best-dev to mert-moses.pl. This will copy the moses.ini file corresponding to the highest-scoring development run (as determined by the evaluator executable using BLEU on run*.out) into the final moses.ini. This can make a fairly big difference for MIRA’s test-time accuracy. You can also pass through options to kbmira by adding --batch-mira-args ’whatever’ to mert-moses.pl. Useful kbmira options include: • -J n : changes the number of inner MIRA loops to n passes over the data. Increasing this value to 100 or 300 can be good for working with small development sets. The default, 60, is ideal for development sets with more than 1000 sentences. • -C n : changes MIRA’s C-value to n. This controls regularization. The default, 0.01, works well for most situations, but if it looks like MIRA is over-fitting or not converging, decreasing C to 0.001 or 0.0001 can sometimes help. • --streaming : stream k-best lists from disk rather than load them into memory. This results in very slow training, but may be necessary in low-memory environments or with very large development sets. Run kbmira --help for a full list of options. So, a complete call might look like this:$MOSES_SCRIPTS/training/mert-moses.pl work/dev.fr work/dev.en \
$MOSES_BIN/moses work/model/moses.ini --mertdir$MOSES_BIN \
--rootdir MOSES_SCRIPTS --batch-mira --return-best-dev \ --batch-mira-args ’-J 300’ --decoder-flags ’-threads 8 -v 0’ Please give it a try. If it’s not working as advertised, send Colin Cherry an e-mail. For more information on batch MIRA, check out the paper: Colin Cherry and George Foster: "Batch Tuning Strategies for Statistical Machine Translation", NAACL, June 2012, pdf34 34 https://sites.google.com/site/colinacherry/Cherry_Foster_NAACL_2012.pdf 248 5. Training Manual Anticipating some questions: [Q: Does it only handle BLEU?] [A: Yes, for now. There’s nothing stopping people from implementing other metrics, so long as a reasonable sentence-level version of the metric can be worked out. Note that you generally need to retune kbmira’s C-value for different metrics. I’d also change --return-best-dev to use the new metric as well.] [Q: Have you tested this on a cluster?] [A: No, I don’t have access to a Sun Grid cluster - I would love it if someone would test that scenario for me. But it works just fine using multi-threaded decoding. Since training happens in a batch, decoding is embarrassingly parallel.] Tree-to-string and tree-to-tree tuning When tuning with tree input, make sure you set the inputtype argument to the mert script mert-moses.pl --inputtype 3 ... Subsection last modified on May 21, 2014, at 11:11 AM 6 Background 6.1 Background Statistical Machine Translation as a research area started in the late 1980s with the Candide project at IBM. IBM’s original approach maps individual words to words and allows for deletion and insertion of words. Lately, various researchers have shown better translation quality with the use of phrase translation. Phrase-based MT can be traced back to Och’s alignment template model, which can be re-framed as a phrase translation system. Other researchers used augmented their systems with phrase translation, such as Yamada, who use phrase translation in a syntax-based model. Marcu introduced a joint-probability model for phrase translation. At this point, most competitive statistical machine translation systems use phrase translation, such as the CMU, IBM, ISI, and Google systems, to name just a few. Phrase-based systems came out ahead at a recent international machine translation competition (DARPA TIDES Machine Translation Evaluation 2003-2006 on Chinese-English and Arabic-English). Of course, there are other ways to do machine translation. Most commercial systems use transfer rules and a rich translation lexicon. Machine translation research was focused on transferbased systems in the 1980s and on knowledge based systems that use an interlingua representation as an intermediate step between input and output in the 1990s. There are also other ways to do statistical machine translation. There is some effort in building syntax-based models that either use real syntax trees generated by syntactic parsers, or tree transfer methods motivated by syntactic reordering patterns. The phrase-based statistical machine translation model we present here was defined by Koehn et al. (2003)1 . See also the description by Zens (2002)2 . The alternative phrase-based methods differ in the way the phrase translation table is created, which we discuss in detail below. 1 2 http://acl.ldc.upenn.edu/N/N03/N03-1017.pdf http://www.rzens.com/Zens_KI_2002.pdf 249 250 6. Background 6.1.1 Model The figure below illustrates the process of phrase-based translation. The input is segmented into a number of sequences of consecutive words (so-called phrases). Each phrase is translated into an English phrase, and English phrases in the output may be reordered. In this section, we will define the phrase-based machine translation model formally. The phrase translation model is based on the noisy channel model. We use Bayes rule to reformulate the translation probability for translating a foreign sentence f into English e as argmaxe p(e|f) = argmaxe p(f|e) p(e) This allows for a language model p(e) and a separate translation model p(f|e). I During decoding, the foreign input sentence f is segmented into a sequence of I phrases f¯1 . We assume a uniform probability distribution over all possible segmentations. I Each foreign phrase f¯i in f¯1 is translated into an English phrase ēi . The English phrases may be reordered. Phrase translation is modeled by a probability distribution φ(f¯i |ēi ). Recall that due to the Bayes rule, the translation direction is inverted from a modeling standpoint. Reordering of the English output phrases is modeled by a relative distortion probability distribution d(starti ,endi-1 ), where starti denotes the start position of the foreign phrase that was translated into the ith English phrase, and endi-1 denotes the end position of the foreign phrase that was translated into the (i-1)th English phrase. We use a simple distortion model d(starti ,endi-1 ) = α|starti -endi-1 -1| with an appropriate value for the parameter α. In order to calibrate the output length, we introduce a factor ω (called word cost) for each generated English word in addition to the trigram language model pLM . This is a simple means to optimize performance. Usually, this factor is larger than 1, biasing toward longer output. In summary, the best English output sentence ebest given a foreign input sentence f according to our model is ebest = argmaxe p(e|f) = argmaxe p(f|e) pLM (e) ω length(e) where p(f|e) is decomposed into I p(f¯1 |ē1 I ) = Φi=1 I φ(f¯i |ēi ) d(starti ,endi-1 ) 6.1. 6.1.2 Background 251 Word Alignment When describing the phrase-based translation model so far, we did not discuss how to obtain the model parameters, especially the phrase probability translation table that maps foreign phrases to English phrases. Most recently published methods on extracting a phrase translation table from a parallel corpus start with a word alignment. Word alignment is an active research topic. For instance, this problem was the focus as a shared task at a recent data driven machine translation workshop3 . See also the systematic comparison by Och and Ney (Computational Linguistics, 2003). At this point, the most common tool to establish a word alignment is to use the toolkit GIZA++4 . This toolkit is an implementation of the original IBM models that started statistical machine translation research. However, these models have some serious draw-backs. Most importantly, they only allow at most one English word to be aligned with each foreign word. To resolve this, some transformations are applied. First, the parallel corpus is aligned bidirectionally, e.g., Spanish to English and English to Spanish. This generates two word alignments that have to be reconciled. If we intersect the two alignments, we get a high-precision alignment of high-confidence alignment points. If we take the union of the two alignments, we get a high-recall alignment with additional alignment points. See the figure below for an illustration. Researchers differ in their methods where to go from here. We describe the details below. 3 4 http://www.statmt.org/wpt05/ http://www.isi.edu/~och/GIZA++.html 252 6.1.3 6. Background Methods for Learning Phrase Translations Most of the recently proposed methods use a word alignment to learn a phrase translation table. We discuss three such methods in this section and one exception. Marcu and Wong First, the exception: Marcu and Wong (EMNLP, 2002) proposed to establish phrase correspondences directly in a parallel corpus. To learn such correspondences, they introduced a phrasebased joint probability model that simultaneously generates both the source and target sentences in a parallel corpus. Expectation Maximization learning in Marcu and Wong’s framework yields both (i) a joint probability distribution φ(ē, f¯), which reflects the probability that phrases ē and f¯ are translation equivalents; (ii) and a joint distribution d(i,j), which reflects the probability that a phrase at position i is translated into a phrase at position j. To use this model in the context of our framework, we simply marginalize the joint probabilities estimated by Marcu and Wong (EMNLP, 2002) to conditional probabilities. Note that this approach is consistent with the approach taken by Marcu and Wong themselves, who use conditional models during decoding. 6.1.4 Och and Ney Och and Ney (Computational Linguistics, 2003) propose a heuristic approach to refine the alignments obtained from GIZA++. At a minimum, all alignment points of the intersection of the two alignments are maintained. At a maximum, the points of the union of the two alignments are considered. To illustrate this, see the figure below. The intersection points are black, the additional points in the union are shaded grey. 6.1. Background 253 Och and Ney explore the space between intersection and union with expansion heuristics that start with the intersection and add additional alignment points. The decision which points to add may depend on a number of criteria: • In which alignment does the potential alignment point exist? Foreign-English or Englishforeign? • Does the potential point neighbor already established points? • Does neighboring mean directly adjacent (block-distance), or also diagonally adjacent? • Is the English or the foreign word that the potential point connects unaligned so far? Are both unaligned? • What is the lexical probability for the potential point? Och and Ney (Computational Linguistics, 2003) are ambigous in their description about which alignment points are added in their refined method. We reimplemented their method for Moses, so we will describe this interpretation here. Our heuristic proceeds as follows: We start with intersection of the two word alignments. We only add new alignment points that exist in the union of two word alignments. We also always require that a new alignment point connects at least one previously unaligned word. First, we expand to only directly adjacent alignment points. We check for potential points starting from the top right corner of the alignment matrix, checking for alignment points for the first English word, then continue with alignment points for the second English word, and so on. This is done iteratively until no alignment point can be added anymore. In a final step, we add non-adjacent alignment points, with otherwise the same requirements. We collect all aligned phrase pairs that are consistent with the word alignment: The words in a legal phrase pair are only aligned to each other, and not to words outside. The set of bilingual phrases BP can be defined formally (Zens, KI 2002) as: BP (f1J , eJ1 , A) = {(fjj+m , ei+n )} : ∀(i0 , j 0 ) ∈ A : j <= j 0 <= j + m ↔ i <= i0 <= i + m i See the figure below for some examples what this means. All alignment points for words that are part of the phrase pair have to be in the phrase alignment box. It is fine to have unaligned words in a phrase alignment, even at the boundary. The figure below displays all the phrase pairs that are collected according to this definition for the alignment from our running example. 254 6. Background (Maria, Mary), (no, did not), (slap, daba una bofetada), (a la, the), (bruja, witch), (verde, green), (Maria no, Mary did not), (no daba una bofetada, did not slap), (daba una bofetada a la, slap the), (bruja verde, green witch), (Maria no daba una bofetada, Mary did not slap), (no daba una bofetada a la, did not slap the), (a la bruja verde, the green witch) (Maria no daba una bofetada a la, Mary did not slap the), (daba una bofetada a la bruja verde, slap the green witch), (no daba una bofetada a la bruja verde, did not slap the green witch), (Maria no daba una bofetada a la bruja verde, Mary did not slap the green witch) Given the collected phrase pairs, we estimate the phrase translation probability distribution by relative frequency: φ(f¯|ē) = count(f¯,ē) Σf¯ count(f¯,ē) No smoothing is performed, although lexical weighting addresses the problem of sparse data. For more details, see our paper on phrase-based translation (Koehn at al, HLT-NAACL 2003). Tillmann Tillmann (EMNLP, 2003) proposes a variation of this method. He starts with phrase alignments based on the intersection of the two GIZA++ alignments and uses points of the union to expand these. See his presentation for details. Venugopal, Zhang, and Vogel Venugopal et al. (ACL 2003) allows also for the collection of phrase pairs that are violated by the word alignment. They introduce a number of scoring methods take consistency with the word alignment, lexical translation probabilities, phrase length, etc. into account. Zhang et al. (2003) proposes a phrase alignment method that is based on word alignments and tries to find a unique segmentation of the sentence pair, as it is done by Marcu and Wong directly. This enables them to estimate joint probability distributions, which can be marginalized into conditional probability distributions. Vogel et al. (2003) reviews these two methods and shows that the combining phrase tables generated by different methods improves results. 6.2. Decoder 6.2 255 Decoder This section describes the Moses decoder from a more theoretical perspective. The decoder was originally developed for the phrase model proposed by Marcu and Wong. At that time, only a greedy hill-climbing decoder was available, which was unsufficent for our work on noun phrase translation (Koehn, PhD, 2003). The decoder implements a beam search and is roughly similar to work by Tillmann (PhD, 2001) and Och (PhD, 2002). In fact, by reframing Och’s alignment template model as a phrase translation model, the decoder is also suitable for his model, as well as other recently proposed phrase models. We start this section with defining the concept of translation options, describe the basic mechanism of beam search, and its neccessary components: pruning, future cost estimates. We conclude with background on n-best list generation. 6.2.1 Translation Options Given an input string of words, a number of phrase translations could be applied. We call each such applicable phrase translation a translation option. This is illustrated in the figure below, where a number of phrase translations for the Spanish input sentence Maria no daba uma bofetada a la bruja verde are given. These translation options are collected before any decoding takes place. This allows a quicker lookup than consulting the whole phrase translation table during decoding. The translation options are stored with the information • • • • first foreign word covered last foreign word covered English phrase translation phrase translation probability Note that only the translation options that can be applied to a given input text are necessary for decoding. Since the entire phrase translation table may be too big to fit into memory, we can restrict ourselves to these translation options to overcome such computational concerns. We may even generate a phrase translation table on demand that only includes valid translation options for a given input text. This way, a full phrase translation table (that may be computationally too expensive to produce) may never have to be built. 256 6. Background 6.2.2 Core Algorithm The phrase-based decoder we developed employs a beam search algorithm, similar to the one used by Jelinek (book "Statistical Methods for Speech Recognition", 1998) for speech recognition. The English output sentence is generated left to right in form of hypotheses. This process illustrated in the figure below. Starting from the initial hypothesis, the first expansion is the foreign word Maria, which is translated as Mary. The foreign word is marked as translated (marked by an asterisk). We may also expand the initial hypothesis by translating the foreign word bruja as witch. We can generate new hypotheses from these expanded hypotheses. Given the first expanded hypothesis we generate a new hypothesis by translating no with did not. Now the first two foreign words Maria and no are marked as being covered. Following the back pointers of the hypotheses we can read of the (partial) translations of the sentence. Let us now describe the beam search more formally. We begin the search in an initial state where no foreign input words are translated and no English output words have been generated. New states are created by extending the English output with a phrasal translation of that covers some of the foreign input words not yet translated. The current cost of the new state is the cost of the original state multiplied with the translation, distortion and language model costs of the added phrasal translation. Note that we use the informal concept cost analogous to probability: A high cost is a low probability. Each search state (hypothesis) is represented by • a back link to the best previous state (needed for finding the best translation of the sentence by back-tracking through the search states) • the foreign words covered so far • the last two English words generated (needed for computing future language model costs) • the end of the last foreign phrase covered (needed for computing future distortion costs) • the last added English phrase (needed for reading the translation from a path of hypotheses) • the cost so far • an estimate of the future cost (is precomputed and stored for efficiency reasons, as detailed in below in special section) 6.2. Decoder 257 Final states in the search are hypotheses that cover all foreign words. Among these the hypothesis with the lowest cost (highest probability) is selected as best translation. The algorithm described so far can be used for exhaustively searching through all possible translations. In the next sections we will describe how to optimize the search by discarding hypotheses that cannot be part of the path to the best translation. We then introduce the concept of comparable states that allow us to define a beam of good hypotheses and prune out hypotheses that fall out of this beam. In a later section, we will describe how to generate an (approximate) n-best list. 6.2.3 Recombining Hypotheses Recombining hypothesis is a risk-free way to reduce the search space. Two hypotheses can be recombined if they agree in • the foreign words covered so far • the last two English words generated • the end of the last foreign phrase covered If there are two paths that lead to two hypotheses that agree in these properties, we keep only the cheaper hypothesis, e.g., the one with the least cost so far. The other hypothesis cannot be part of the path to the best translation, and we can safely discard it. Note that the inferior hypothesis can be part of the path to the second best translation. This is important for generating n-best lists. 6.2.4 Beam Search While the recombination of hypotheses as described above reduces the size of the search space, this is not enough for all but the shortest sentences. Let us estimate how many hypotheses (or, states) are generated during an exhaustive search. Considering the possible values for the properties of unique hypotheses, we can estimate an upper bound for the number of states by N ~ 2nf |Ve |2 nf where nf is the number of foreign words, and |Ve | the size of the English vocabulary. In practice, the number of possible English words for the last two words generated is much smaller than |Ve |2 . The main concern is the exponential explosion from the 2nf possible configurations of foreign words covered by a hypothesis. Note this causes the problem of machine translation to become NP-complete (Knight, Computational Linguistics, 1999) and thus dramatically harder than, for instance, speech recognition. In our beam search we compare the hypotheses that cover the same number of foreign words and prune out the inferior hypotheses. We could base the judgment of what inferior hypotheses are on the cost of each hypothesis so far. However, this is generally a very bad criterion, since it biases the search to first translating the easy part of the sentence. For instance, if there is a three word foreign phrase that easily translates into a common English phrase, this may carry much less cost than translating three words separately into uncommon English words. The search will prefer to start the sentence with the easy part and discount alternatives too early. So, our measure for pruning out hypotheses in our beam search does not only include the cost so far, but also an estimate of the future cost. This future cost estimation should favor hypotheses that already covered difficult parts of the sentence and have only easy parts left, 258 6. Background and discount hypotheses that covered the easy parts first. We describe the details of our future cost estimation in the next section. Given the cost so far and the future cost estimation, we can prune out hypotheses that fall outside the beam. The beam size can be defined by threshold and histogram pruning. A relative threshold cuts out a hypothesis with a probability less than a factor α of the best hypotheses (e.g., α = 0.001). Histogram pruning keeps a certain number n of hypotheses (e.g., n = 100). Note that this type of pruning is not risk-free (opposed to the recombination, which we described earlier). If the future cost estimates are inadequate, we may prune out hypotheses on the path to the best scoring translation. In a particular version of beam search, A* search, the future cost estimate is required to be _admissible_, which means that it never overestimates the future cost. Using best-first search and an admissible heuristic allows pruning that is risk-free. In practice, however, this type of pruning does not sufficiently reduce the search space. See more on search in any good Artificial Intelligence text book, such as the one by Russel and Norvig ("Artificial Intelligence: A Modern Approach"). The figure below gives pseudo-code for the algorithm we used for our beam search. For each number of foreign words covered, a hypothesis stack in created. The initial hypothesis is placed in the stack for hypotheses with no foreign words covered. Starting with this hypothesis, new hypotheses are generated by committing to phrasal translations that covered previously unused foreign words. Each derived hypothesis is placed in a stack based on the number of foreign words it covers. initialize hypothesisStack[0 .. nf]; create initial hypothesis hyp_init; add to stack hypothesisStack[0]; for i=0 to nf-1: for each hyp in hypothesisStack[i]: for each new_hyp that can be derived from hyp: nf[new_hyp] = number of foreign words covered by new_hyp; add new_hyp to hypothesisStack[nf[new_hyp]]; prune hypothesisStack[nf[new_hyp]]; find best hypothesis best_hyp in hypothesisStack[nf]; output best path that leads to best_hyp; We proceed through these hypothesis stacks, going through each hypothesis in the stack, deriving new hypotheses for this hypothesis and placing them into the appropriate stack (see figure below for an illustration). After a new hypothesis is placed into a stack, the stack may have to be pruned by threshold or histogram pruning, if it has become too large. In the end, the best hypothesis of the ones that cover all foreign words is the final state of the best translation. We can read off the English words of the translation by following the back links in each hypothesis. 6.2. Decoder 6.2.5 259 Future Cost Estimation Recall that for excluding hypotheses from the beam we do not only have to consider the cost so far, but also an estimate of the future cost. While it is possible to calculate the cheapest possible future cost for each hypothesis, this is computationally so expensive that it would defeat the purpose of the beam search. The future cost is tied to the foreign words that are not yet translated. In the framework of the phrase-based model, not only may single words be translated individually, but also consecutive sequences of words as a phrase. Each such translation operation carries a translation cost, language model costs, and a distortion cost. For our future cost estimate we consider only translation and language model costs. The language model cost is usually calculated by a trigram language model. However, we do not know the preceding English words for a translation operation. Therefore, we approximate this cost by computing the language model score for the generated English words alone. That means, if only one English word is generated, we take its unigram probability. If two words are generated, we take the unigram probability of the first word and the bigram probability of the second word, and so on. For a sequence of foreign words multiple overlapping translation options exist. We just described how we calculate the cost for each translation option. The cheapest way to translate the sequence of foreign words includes the cheapest translation options. We approximate the cost for a path through translation options by the product of the cost for each option. To illustrate this concept, refer to the figure below. The translation options cover different consecutive foreign words and carry an estimated cost cij . The cost of the shaded path through the sequence of translation options is c01 c12 c25 = 1.9578 * 10-7 . 260 6. Background The cheapest path for a sequence of foreign words can be quickly computed with dynamic programming. Also note that if the foreign words not covered so far are two (or more) disconnected sequences of foreign words, the combined cost is simply the product of the costs for each contiguous sequence. Since there are only n(n+1)/2 contiguous sequences for n words, the future cost estimates for these sequences can be easily precomputed and cached for each input sentence. Looking up the future costs for a hypothesis can then be done very quickly by table lookup. This has considerable speed advantages over computing future cost on the fly. 6.2.6 N-Best Lists Generation Usually, we expect the decoder to give us the best translation for a given input according to the model. But for some applications, we might also be interested in the second best translation, third best translation, and so on. A common method in speech recognition, that has also emerged in machine translation is to first use a machine translation system such as our decoder as a base model to generate a set of candidate translations for each input sentence. Then, additional features are used to rescore these translations. An n-best list is one way to represent multiple candidate translations. Such a set of possible translations can also be represented by word graphs (Ueffing et al., EMNLP 2002) or forest structures over parse trees (Langkilde, EACL 2002). These alternative data structures allow for more compact representation of a much larger set of candidates. However, it is much harder to detect and score global properties over such data structures. Additional Arcs in the Search Graph Recall the process of state expansions. The generated hypotheses and the expansions that link them form a graph. Paths branch out when there are multiple translation options for a hypothesis from which multiple new hypotheses can be derived. Paths join when hypotheses are recombined. Usually, when we recombine hypotheses, we simply discard the worse hypothesis, since it cannot possibly be part of the best path through the search graph (in other words, part of the best translation). But since we are now also interested in the second best translation, we cannot simply discard information about that hypothesis. If we would do this, the search graph would only contain one path for each hypothesis in the last hypothesis stack (which contains hypotheses that cover all foreign words). If we store information that there are multiple ways to reach a hypothesis, the number of possible paths also multiplies along the path when we traverse backward through the graph. In order to keep the information about merging paths, we keep a record of such merges that contains • identifier of the previous hypothesis • identifier of the lower-cost hypothesis • cost from the previous to higher-cost hypothesis 6.2. Decoder 261 The figure below gives an example for the generation of such an arc: in this case, the hypotheses 2 and 4 are equivalent in respect to the heuristic search, as detailed above. Hence, hypothesis 4 is deleted. But since we want to keep the information about the path leading from hypothesis 3 to 2, we store a record of this arc. The arc also contains the cost added from hypothesis 3 to 4. Note that the cost from hypothesis 1 to hypothesis 2 does not have to be stored, since it can be recomputed from the hypothesis data structures. Mining the Search Graph for an n-Best List The graph of the hypothesis space can be also be viewed as a probabilistic finite state automaton. The hypotheses are states, and the records of back-links and the additionally stored arcs are state transitions. The added probability scores when expanding a hypothesis are the costs of the state transitions. Finding the n-best path in such a probabilistic finite state automaton is a well-studied problem. In our implementation, we store the information about hypotheses, hypothesis transitions, and additional arcs in a file that can be processed by the finite state toolkit Carmel5 , which we use to mine the n-best lists. This toolkit uses the _n_ shortest paths algorithm by Eppstein (FOCS, 1994). Our method is related to work by Ueffing (2002) for generating n-best lists for IBM Model 4. Subsection last modified on July 28, 2013, at 09:56 AM 5 http://www.isi.edu/licensed-sw/carmel/ 262 6. Background 6.3 Factored Translation Models The current state-of-the-art approach to statistical machine translation, so-called phrase-based models, are limited to the mapping of small text chunks (phrases) without any explicit use of linguistic information, may it be morphological, syntactic, or semantic. Such additional information has been demonstrated to be valuable by integrating it in pre-processing or postprocessing. However, a tighter integration of linguistic information into the translation model is desirable for two reasons: • Translation models that operate on more general representations, such as lemmas instead of surface forms of words, can draw on richer statistics and overcome the data sparseness problems caused by limited training data. • Many aspects of translation can be best explained on a morphological, syntactic, or semantic level. Having such information available to the translation model allows the direct modeling of these aspects. For instance: reordering at the sentence level is mostly driven by general syntactic principles, local agreement constraints show up in morphology, etc. Therefore, we developed a framework for statistical translation models that tightly integrates additional information. Our framework is an extension of the phrase-based approach. It adds additional annotation at the word level. A word in our framework is not anymore only a token, but a vector of factors that represent different levels of annotation (see figure below). 6.3.1 Motivating Example: Morphology One example to illustrate the short-comings of the traditional surface word approach in statistical machine translation is the poor handling of morphology. Each word form is treated as a token in itself. This means that the translation model treats, say, the word house completely independent of the word houses. Any instance of house in the training data does not add any knowledge to the translation of houses. In the extreme case, while the translation of house may be known to the model, the word houses may be unknown and the system will not be able to translate it. While this problem does not 6.3. Factored Translation Models 263 show up as strongly in English - due to the very limited morphological production in English - it does constitute a significant problem for morphologically rich languages such as Arabic, German, Czech, etc. Thus, it may be preferably to model translation between morphologically rich languages on the level of lemmas, and thus pooling the evidence for different word forms that derive from a common lemma. In such a model, we would want to translate lemma and morphological information separately, and combine this information on the output side to ultimately generate the output surface words. Such a model can be defined straight-forward as a factored translation model. See figure below for an illustration of this model in our framework. Note that while we illustrate the use of factored translation models on such a linguistically motivated example, our framework also applies to models that incorporate statistically defined word classes, or any other annotation. 6.3.2 Decomposition of Factored Translation The translation of factored representations of input words into the factored representations of output words is broken up into a sequence of mapping steps that either translate input factors into output factors, or generate additional output factors from existing output factors. Recall the example of a factored model motivated by morphological analysis and generation. In this model the translation process is broken up into the following three mapping steps: • Translate input lemmas into output lemmas • Translate morphological and POS factors • Generate surface forms given the lemma and linguistic factors Factored translation models build on the phrase-based approach, which defines a segmentation of the input and output sentences into phrases. Our current implementation of factored translation models follows strictly the phrase-based approach, with the additional decomposition of phrase translation into a sequence of mapping steps. Since all mapping steps operate on the same phrase segmentation of the input and output sentence into phrase pairs, we call these synchronous factored models. 264 6. Background Let us now take a closer look at one example, the translation of the one-word phrase h"auser into English. The representation of h"auser in German is: surface-form h"auser | lemma haus | part-of-speech NN | count plural | case nominative | gender neutral. The three mapping steps in our morphological analysis and generation model may provide the following applicable mappings: • Translation: Mapping lemmas – haus -> house, home, building, shell • Translation: Mapping morphology – NN|plural-nominative-neutral -> NN|plural, NN|singular • Generation: Generating surface forms – house|NN|plural -> houses – house|NN|singular -> house – home|NN|plural -> homes – ... We call the application of these mapping steps to an input phrase expansion. Given the multiple choices for each step (reflecting the ambiguity in translation), each input phrase may be expanded into a list of translation options. The German h"auser|haus|NN|plural-nominative-neutral may be expanded as follows: • Translation: Mapping lemmas – { ?|house|?|?, ?|home|?|?, ?|building|?|?, ?|shell|?|? } • Translation: Mapping morphology – { ?|house|NN|plural, ?|home|NN|plural, ?|building|NN|plural, ?|shell|NN|plural, ?|house|NN|singular,... } • Generation: Generating surface forms – { houses|house|NN|plural, homes|home|NN|plural, buildings|building|NN|plural, shells|shell|NN|plural, house|house|NN|singular, ... } 6.3.3 Statistical Model Factored translation models follow closely the statistical modeling approach of phrase-based models (in fact, phrase-based models are a special case of factored models). The main difference lies in the preparation of the training data and the type of models learned from the data. Training The training data (a parallel corpus) has to be annotated with the additional factors. For instance, if we want to add part-of-speech information on the input and output side, we need to obtain part-of-speech tagged training data. Typically this involves running automatic tools on the corpus, since manually annotated corpora are rare and expensive to produce. Next, we need to establish a word-alignment for all the sentences in the parallel training corpus. Here, we use the same methodology as in phrase-based models (symmetrized GIZA++ alignments). The word alignment methods may operate on the surface forms of words, or on any of the other factors. In fact, some preliminary experiments have shown that word alignment based on lemmas or stems yields improved alignment quality. 6.3. Factored Translation Models 265 Each mapping step forms a component of the overall model. From a training point of view this means that we need to learn translation and generation tables from the word-aligned parallel corpus and define scoring methods that help us to choose between ambiguous mappings. Phrase-based translation models are acquired from a word-aligned parallel corpus by extracting all phrase-pairs that are consistent with the word alignment. Given the set of extracted phrase pairs with counts, various scoring functions are estimated, such as conditional phrase translation probabilities based on relative frequency estimation or lexical translation probabilities based on the words in the phrases. In our approach, the models for the translation steps are acquired in the same manner from a word-aligned parallel corpus. For the specified factors in the input and output, phrase mappings are extracted. The set of phrase mappings (now over factored representations) is scored based on relative counts and word-based translation probabilities. The tables for generation steps are estimated on the output side only. The word alignment plays no role here. In fact, additional monolingual data may be used. The generation model is learned on a word-for-word basis. For instance, for a generation step that maps surface forms to part-of-speech, a table with entries such as (fish,NN) is constructed. One or more scoring functions may be defined over this table, in our experiments we used both conditional probability distributions, e.g., p(fish|NN) and p(NN|fish), obtained by maximum likelihood estimation. An important component of statistical machine translation is the language model, typically an n-gram model over surface forms of words. In the framework of factored translation models, such sequence models may be defined over any factor, or any set of factors. For factors such as part-of-speech tags, building and using higher order n-gram models (7-gram, 9-gram) is straight-forward. Combination of Components As in phrase-based models, factored translation models can be seen as the combination of several components (language model, reordering model, translation steps, generation steps). These components define one or more feature functions that are combined in a log-linear model: e|f ) = exp n X λi hi (e, f) i=1 To compute the probability of a translation e given an input sentence f, we have to evaluate each feature function hi . For instance, the feature function for a bigram language model component is (m is the number of words ei in the sentence e): hlm (e, f) = plm (e) = p(e1 )p(e2 |p1 )...p(em |em−1 ) Let us now consider the feature functions introduced by the translation and generation steps of factored translation models. The translation of the input sentence f into the output sentence e breaks down to a set of phrase translations (f¯j , e¯j ). For a translation step component, each feature function ht is defined over the phrase pairs (f¯j ,ēj ) given a scoring function τ : 266 6. Background ht (e, f) = Σj τ (f¯j , ēj )00 For a generation step component, each feature function hg given a scoring function γ is defined over the output words ek only: hg (e, f) = Σk γ(ek ) The feature functions follow from the scoring functions (τ , γ) acquired during the training of translation and generation tables. For instance, recall our earlier example: a scoring function for a generation model component that is a conditional probability distribution between input and output factors, e.g., γ(fish,NN,singular) = p(NN|fish). The feature weights λi in the log-linear model are determined with the usual minimum error rate training method. Efficient Decoding Compared to phrase-based models, the decomposition of phrase translation into several mapping steps creates additional computational complexity. Instead of a simple table lookup to obtain the possible translations for an input phrase, now multiple tables have to be consulted and their content combined. In phrase-based models it is easy to identify the entries in the phrase table that may be used for a specific input sentence. These are called translation options. We usually limit ourselves to the top 20 translation options for each input phrase. The beam search decoding algorithm starts with an empty hypothesis. Then new hypotheses are generated by using all applicable translation options. These hypotheses are used to generate further hypotheses in the same manner, and so on, until hypotheses are created that cover the full input sentence. The highest scoring complete hypothesis indicates the best translation according to the model. How do we adapt this algorithm for factored translation models? Since all mapping steps operate on the same segmentation, the expansions of these mapping steps can be efficiently pre-computed prior to the heuristic beam search, and stored as translation options. For a given input phrase, all possible translation options are thus computed before decoding (recall the earlier example, where we carried out the expansion for one input phrase). This means that the fundamental search algorithm does not change. However, we need to be careful about combinatorial explosion of the number of translation options given a sequence of mapping steps. In other words, the expansion may create too many translation options to handle. If one or many mapping steps result in a vast increase of (intermediate) expansions, this may be become unmanageable. We currently address this problem by early pruning of expansions, and limiting the number of translation options per input phrase to a maximum number, by default 50. This is, however, not a perfect solution. Subsection last modified on December 22, 2008, at 06:20 AM 6.4. 6.4 Confusion Networks Decoding 267 Confusion Networks Decoding Machine translation input currently takes the form of simple sequences of words. However, there are increasing demands to integrate machine translation technology in larger information processing systems with upstream natural language and/or speech processing tools (such as named entity recognizers, automatic speech recognizers, morphological analyzers, etc.). These upstream processes tend to generate multiple, erroneous hypotheses with varying confidence. Current MT systems are designed to process only one input hypothesis, making them vulnerable to errors in the input. We extend current MT decoding methods to process multiple, ambiguous hypotheses in the form of an input lattice. A lattice representation allows an MT system to arbitrate between multiple ambiguous hypotheses from upstream processing so that the best translation can be produced. As lattice has usually a complex topology, an approximation of it, called confusion network, is used instead. The extraction of a confusion network from a lattice can be performed by means of a publicly available lattice-tool contained in the SRILM toolkit. See the SRILM manual pages6 for details and user guide. 6.4.1 Confusion Networks A Confusion Network (CN), also known as a sausage, is a weighted directed graph with the peculiarity that each path from the start node to the end node goes through all the other nodes. Each edge is labeled with a word and a (posterior) probability. The total probability of all edges between two consecutive nodes sum up to 1. Notice that this is not a strict constraint from the point of view of the decoder; any score can be provided. A path from the start node to the end node is scored by multiplying the scores of its edges. If the previous constrain is satisfied, the product represents the likelihood of the path, and the sum of the likelihood of all paths equals to 1. Between any two consecutive nodes, one (at most) special word _eps_ can be inserted; _eps_ words allows paths having different lengths. Any path within a CN represents a realization of the CN. Realizations of a CN can differ in terms of either sequence of words or total score. It is possible that two (or more) realizations have the same sequence of words, but different scores. Word lengths can also differ due to presence of the _eps_. This is a list of some realization of the previous CN. 6 http://www.speech.sri.com/projects/srilm/manpages 268 6. Background aus der Zeitung Aus der Zeitung Zeitung Haus Zeitungs score=0.252 score=0.126 score=0.021 score=0.001 length=3 length=3 length=1 length=2 Notes • A CN contains all paths of the lattice which is originated from. • A CN can contain more paths than the lattice which is originated from (due _eps_). 6.4.2 Representation of Confusion Network Moses adopts the following computer-friendly representation for a CN. Haus 0.1 aus 0.4 _eps_ 0.3 Aus 0.2 der 0.9 _eps_ 0.1 Zeitung 0.7 _eps_ 0.2 Zeitungs 0.1 where a line contains the alternative edges (words and probs) between two consecutive nodes. In the factored representation, each line gives alternatives over the full factor space: Haus|N 0.1 aus|PREP 0.4 Aus|N 0.4 _eps_|_eps_ 0.1 der|DET 0.1 der|PREP 0.8 _eps_|_eps_ 0.1 Zeitung|N 0.7 _eps_|_eps_ 0.2 Zeitungs|N 0.1 Notice that if you project the above CN on a single factor, repetitions of factors must be merged and the respective probs summed up. The corresponding word-projected CN is the one of the first example, while the part-of-speech projected CN is: N 0.5 PREP 0.4 _eps_ 0.1 DET 0.1 PREP 0.8 _eps_ 0.1 N 0.8 _eps_ 0.2 Subsection last modified on March 26, 2013, at 10:18 AM 6.5. 6.5 Word Lattices 269 Word Lattices A word lattice is a directed acyclic graph with a single start point and edges labeled with a word and weight. Unlike confusion networks which additionally impose the requirement that every path must pass through every node, word lattices can represent any finite set of strings (although this generality makes word lattices slightly less space-efficient than confusion networks). However, in general a word lattice can represent an exponential number of sentences in polynomial space. Here is an example lattice showing possible ways of decompounding some compound words in German: Moses can decode input represented as a word lattice, and, in most useful cases, do this far more efficiently than if each sentence encoded in the lattice were decoded serially. When Moses translates input encoded as a word lattice the translation it chooses maximizes the translation probability along any path in the input (but, to be clear, a single translation hypothesis in Moses corresponds to a single path through the input lattice). 6.5.1 How to represent lattice inputs Lattices are encoded by ordering the nodes in a topological ordering (there may be more than one way to do this- in general, any one is as good as any other, but see the comments on -max-phrase-length below) and using this ordering to assign consecutive numerical IDs to the nodes. Then, proceeding in order through the nodes, each node lists its outgoing edges and any weights associated with them. For example, the above lattice can be written in the moses format (also called the Python lattice format -- PLF): ( ( (’einen’, 1.0, 1), ), ( (’wettbewerbsbedingten’, 0.5, 2), (’wettbewerbs’, 0.25, 1), (’wettbewerb’, 0.25, 1), ), ( (’bedingten’, 1.0, 1), ), ( (’preissturz’, 0.5, 2), (’preis’, 0.5, 1), ), 270 6. Background ( (’sturz’, 1.0, 1), ), ) The second number is the probability associated with an edge. The third number is distance between the start and end nodes of the edge, defined as the numerical ID of the end node minus the numerical ID of the start node. Note that the nodes must be numbered in topological order for the distance calculation. Typically, one writes lattices this with no spaces, on a single line as follows: (((’einen’,1.0,1),),((’wettbewerbsbedingten’,0.5,2),(’wettbewerbs’,0.25,1), \ (’wettbewerb’,0.25, 1),),((’bedingten’,1.0,1),),((’preissturz’,0.5,2), \ (’preis’,0.5,1),),((’sturz’,1.0,1),),) 6.5.2 Configuring moses to translate lattices To indicate that moses will be reading lattices in PLF format, you need to specify -inputtype 2 on the command line or in the moses.ini configuration file. Additionally, it is necessary to specify the feature weight that will be used to incorporate arc probability (may not necessarily be a probability!) into the translation model. To do this, add -weight-i X where X is any real number. In word lattices, the phrase length limit imposed by the -max-phrase-length parameter (default: 20) limits the difference between the indices of the start and the end node of a phrase. If your lattice contains long jumps, you may need to increase -max-phrase-length and/or renumber the nodes to make the jumps smaller. 6.5.3 Verifying PLF files with checkplf The command moses-cmd/src/checkplf reads a PLF (lattice format) input file and verifies the format as well as producing statistics. Here’s an example running the application on buggy input: ./checkplf < tanaka.plf Reading PLF from STDIN... Line 1: edge goes beyond goal node at column position 8, edge label = ’&#65332;&#65313;&#65326;&#65313;&#65323;&#65313;’ Goal node expected at position 12, but edge references a node at position 13 Here’s an example running the application on good input: 6.6. Publications 271 christopher-dyers-macbook:src redpony ./checkplf < ok.plf
PLF format appears to be correct.
STATISTICS:
Number of lattices: 1
Total number of nodes: 7
Total number of edges: 9
Average density: 1.28571 edges/node
Total number of paths: 4
Average number of paths: 4
6.5.4
Citation
If you use Moses’s lattice translation in your research, please cite the following paper:
Chris Dyer, Smaranda Muresan, and Philip Resnik. Generalizing Word Lattice Translation7 . In
Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL), July 2008.
6.6
Publications
If you use Moses for your research, please cite the following paper in you publications:
• Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico,
Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, Chris Dyer,
Ondrej Bojar, Alexandra Constantin, Evan Herbst, Moses: Open Source Toolkit for Statistical Machine Translation, Annual Meeting of the Association for Computational Linguistics (ACL), demonstration session, Prague, Czech Republic, June 2007.
You can find out more on how Moses works from the following papers:
• Philipp Koehn and Hieu Hoang. Factored Translation Models, Conference on Empirical
Methods in Natural Language Processing (EMNLP), Prague, Czech Republic, June 2007.
• Richard Zens and Hermann Ney. Efficient Phrase-table Representation for Machine
Translation with Applications to Online MT and Speech Translation, Proceedings of
the Human Language Technology Conference of the North American Chapter of the Association for Computational Linguistics (HLT-NAACL), Rochester, NY, April 2007.
• Nicola Bertoldi, Richard Zens, Marcello Federico and Wade Shen, Efficient Speech Translation through Confusion Network Decoding, IEEE Transactions on Audio, Speech, and
Language Processing, vol. 16, no. 9, pp. 1696-1705, 2008
• Philipp Koehn An Experimental Management System, Proceedings of the Machine Translation Marathon 2010, The Prague Bulletin of Mathematical Linguistics, vol. 94, pp. 86-96,
2010.
7
http://aclweb.org/anthology-new/P/P08/P08-1115.pdf
272
6. Background
• Kenneth Heafield: KenLM: Faster and Smaller Language Model Queries, Proceedings
of the Sixth Workshop on Statistical Machine Translation (WMT), 2011.
• Marcin Junczys-Dowmunt: Phrasal Rank-Encoding: Exploiting Phrase Redundancy
and Translational Relations for Phrase Table Compression, Proceedings of the Machine
Translation Marathon 2012, The Prague Bulletin of Mathematical Linguistics, vol. 98, pp.
63-74, 2012.
• Nadir Durrani, Alexander Fraser, Helmut Schmid, Hieu Hoang, Philipp Koehn. Can
Markov Models Over Minimal Translation Units Help Phrase-Based SMT?, Proceedings of the 51st Annual Conference of the Association for Computational Linguistics
(ACL), Sofia, Bulgaria, August.
• Nadir Durrani, Hassan Sajjad, Hieu Hoang, Philipp Koehn. Integrating an Unsupervised Transliteration Model into Statistical Machine Translation, Proceedings of the
14th Conference of the European Chapter of the Association for Computational Linguistics (EACL), Gothenburg, Sweden, April 2014
7
Code Guide
7.1
Code Guide
7.1.1
Github, branching, and merging
If you want to code with Moses, you should create your own repository in one of a number of
ways.
The preference is that you fork the repository if you’re doing long-term research. If you fixed a
bug, please commit it yourself, or create a pull request.
• Clone the moses github repository to your hard disk and work with it:
git clone https://github.com/moses-smt/mosesdecoder.git mymoses
cd mymoses
edit files ....
git commit -am "Check in"
You don’t need a github login or permission to do this. All changes are stored on your own
hard disk
• Clone AND branch the repository:
git clone https://github.com/moses-smt/mosesdecoder.git mymoses
cd mymoses
git checkout -b mybranch
edit files ....
git commit -am "Check in"
You still don’t need a github login or permission to do this.
273
274
7. Code Guide
• Clone and branch AND push to github:
git clone https://github.com/moses-smt/mosesdecoder.git mymoses
cd mymoses
git checkout -b mybranch
edit files ....
git commit -am "Check in"
git push origin mybranch
edit files ....
git commit -am "Check in again"
git push
You need a github account. And you have to ask one of the Moses administrators to add you
as a committer to the Moses repository.
NB. To delete a LOCAL branch:
git branch -D new-branch
To delete a branch on the github server:
git push origin --delete new-branch
• Fork the repository. You need a github account. You don’t need permission from the
https://github.com/moses-smt/mosesdecoder
Press the Fork button. This creates a new repository only you have write access to. Clone that
repository and do whatever you want. Eg.
git clone https://github.com/hieuhoang/mosesdecoder.git
cd hieuhoang
edit files ....
git commit -am "Check in again"
git push
7.1.
Code Guide
275
• Clone and check into master
git clone https://github.com/moses-smt/mosesdecoder.git mymoses
cd mymoses
edit files ....
git commit -am "Check in"
git push
You need a github account and write permission to the Moses repository.
• Create pull request. Fork a repository and read the instructions here:
https://help.github.com/articles/using-pull-requests
Working with multiple branches
Assuming you’ve done Fork the repository, you can merge the latest changes from the main
Moses repository with this command:
git pull https://github.com/moses-smt/mosesdecoder.git
In your own repository, you can create branches and switch between them using
git checkout -b new-branch
edit files...
git commit -am "check in"
git checkout master
edit files...
...
To get the latest changes from the main Moses repository to your branch, on your fork:
git
git
git
git
checkout master
pull https://github.com/moses-smt/mosesdecoder.git
checkout new-branch
merge master
276
7. Code Guide
Regression test
If you’ve changed any of the C++ code and intend to check into the main Moses repository,
please run the regression test to make sure you haven’t broken anything:
git submodule init
git submodule update
./bjam with-irstlm=... --with-srilm=... {\bf -a --with-regtest} >& reg.out &
Check the output for any failures:
grep FAIL reg.out
Contact Hieu Hoang or Barry Haddow, or any of the other administrators you might know, if
you need help or permission to the github repository.
7.1.2
The code
This section gives a overview of the code. All the source code is commented for Doxygen,
so you can browse a current snapshot of the source documentation1 . Moses is implemented
using object-oriented principles, and you can get a good idea of its class organization from this
documentation
The source code is in the following directories
• moses/util contains some shared utilities, such as code for reading and parsing files, and
for hash tables. This was originally part of KenLM.
• moses/lm contains KenLM, Moses default language model.
• moses/src contains the code for the decoder
• moses/src/LM contains the language model wrappers used by the decoder
• Other subdirectories of moses/src contain some more specialised parts of the decoder,
such as alternative chart decoding algorithms.
• moses-cmd/src contains code relevant to the command line version of the phrase-based
decoder
• moses-cmd/src contains code relevant to the command line version of the chart-based
decoder
• mert contains the code for the Moses mert implementation, originally described here2
In the following, we provide a short walk-through of the decoder.
1
2
http://www.statmt.org/moses/html/hierarchy.html
7.2.
Coding Style
7.1.3
277
Quick Start
• The main function: moses-cmd/src/Main.cpp
• Initialize the decoder
– moses/src/Parameter.cpp specifies parameters
– moses/src/StaticData.cpp contains globals, loads tables
• Process a sentence
– Manager.cpp implements the decoding algorithm
– TranslationOptionCollection.cpp contains translation options
– Hypothesis.cpp represents partial translation
– HypothesisStack.cpp contain viable hypotheses, implements pruning
• Output results: moses-cmd/src/Main.cpp
– moses-cmd/src/IOStream::OutputBestHypo print best translation
– n-best lists generated in Manager.cpp, output in IOStream.cpp
7.1.4
•
•
•
•
•
Detailed Guides
Style Guide (Section 7.2)
Background on factor, word, and phrase data structure (Section 7.3)
Chart decoder (Section 7.4)
7.2
Coding Style
developing Moses.
7.2.1
Formatting
Indentations are 2 spaces. No tab characters allowed in the code. To ensure that your code
follows this format run scripts/other/beautify.perl in the directory of the source code.
Opening braces are on a separate line, for instance:
if (expr) // correct
{
...
}
if (expr) { // wrong
...
}
278
7. Code Guide
Upper/lowercase: Start all functions and class names with capital letters. Start all variable with
small letter. Start all class variable with m_. For instance:
void CalcNBest(size_t count, LatticePathList &ret) const;
Sentence m_source;
Use long variable names, not variables called s, q, or qb7.
Do not use Hungarian notation3 .
7.2.2
The code will be parsed by Doxygen to create online documentation4 . To support this, you have
the Doxygen5 web site.
Class definitions in the *.h file need to be preceeded by a block starting with /**, for instance:
/** The Manager class implements a stack decoding algorithm.
* Hypotheses are organized in stacks. One stack contains all hypothesis that have
* the same number of foreign words translated. The data structure for hypothesis
* stacks is the class HypothesisStack. The data structure for a hypothesis
* is the class Hypothesis.
[...]
**/
class Manager
Class member variable definitions in *.h must be followed by a comment that starts with //!,
for instance:
size_t m_maxNumFactors;
//! max number of factors on both source and target sides
Functions in the *.cpp file need to be preceeded by a block starting with /**, for instance:
3
http://en.wikipedia.org/wiki/Hungarian_notation
http://www.statmt.org/moses/html/hierarchy.html
5
http://www.stack.nl/~dimitri/doxygen/docblocks.html
4
7.2.
Coding Style
279
/**
* Main decoder loop that translates a sentence by expanding
* hypotheses stack by stack, until the end of the sentence.
*/
void Manager::ProcessSentence()
Function parameters are described by param, for instance:
/** Create translation options that exactly cover a specific input span.
* Called by CreateTranslationOptions() and ProcessUnknownWord()
* \param decodeGraph list of decoding steps
* \param factorCollection input sentence with all factors
* \param startPos first position in input sentence
* \param lastPos last position in input sentence
* \param adhereTableLimit whether phrase & generation table limits are adhered to
*/
void TranslationOptionCollection::CreateTranslationOptionsForRange(
const DecodeGraph &decodeGraph
, size_t startPos
, size_t endPos
{
In addition the definition in the *.h must be preceded by a short comment that starts with //!.
This comment will be displayed in the beginning of the class definition. For instance:
//! load all language models as specified in ini file
7.2.3
Data types and methods
• Code for cross-platform compatibility.
• Use object-orientated designs, including
– Create Get/Set functions rather than exposing class variables.
– Label functions, variables and arguments as const where possible.
– Prefer references over pointers
– General styles
– Prefer enum types over integers
– Resolve compiler warnings as well as errors
– Delete tracing/debugging code once they are not needed.
280
7. Code Guide
7.2.4
•
•
•
•
•
Source Control Etiquette
Do not check in non-compilable code, or if functionality is reduced
Ignore the above if you need to, just let people know
Check-in your work often to avoid resolution conflicts
Check in make/project files. However, you are not required to update project files other
than the ones you use.
7.3
Factors, Words, Phrases
Moses is the implemented of a factored translation model. This means that each word is represented by a vector of factors, which are typically word, part-of-speech tags, etc. It also means
that the implementation is a bit more complicated than a non-factored translation model.
This section intends to provide some documentation of how factors, words, and phrases are
implemented in Moses.
7.3.1
Factors
The class Factor6 implements the most basic unit of representing text in Moses. In essence it is
a string.
Factors do not know about their own type (which component in the word vector they represent), this is referred to as its FactorType when needed. This factor type is implemented as a
size_t, i.e. an integer. What a factor really represents (be it a surface form or a part of speech
tag), does not concern the decoder at all. All the decoder knows is that there are a number of
factors that are referred to by their factor type, i.e. an integer index.
Since we do not want to store the same strings over and over again, the class FactorCollection7
contains all known factors. The class has one global instance, and it provides the essential
functions to check if a newly constructed factor already exists and to add a factor. This enables
the comparison of factors by the cheaper comparison of the pointers to factors. Think of the
FactorCollection as the global factor dictionary.
7.3.2
Words
A word is, as we said, a vector of factors. The class Word8 implements this. As data structure,
it is a array over pointers to factors. This does require the code to know what the array size is,
which is set by the global MAX_NUM_FACTORS. The word class implements a number of functions
for comparing and copying words, and the addressing of individual factors.
Again, a word does not know, how many factors it really has. So, for instance, when you want
to print out a word with all its factors, you need to provide also the factor types that are valid
within the word. See the function Word::GetString for details.
6
http://www.statmt.org/moses/html/classMoses_1_1Factor.html
http://www.statmt.org/moses/html/d2/d9f/classMoses_1_1FactorCollection.html
8
http://www.statmt.org/moses/html/d7/dc4/classMoses_1_1Word.html
7
7.4.
Tree-Based Model Decoding
7.3.3
281
Factor Types
This is a good place to note that referring to words gets a bit more complicated. If more than one
factor is used, it does not mean that all the words in the models have all the factors. Take again
the example of a two-factored representation of words as surface form and part-of-speech. We
may still use a simple surface word language model, so for that language model, a word only
has one factor.
We expect the input to the decoder to have all factors specified and during decoding the output
will have all factors of all words set. The process may not be a straight-forward mapping of
the input word to the output word, but it may be decomposed into several mapping steps that
either translate input factors into output factors, or generate additional output factors from
existing output factors.
At this point, keep on mind that a Factor has a FactorType and a Word has a vector<FactorType>,
but these are not internally stored with the Factor and the Word.
Related to factor types is the class FactorMask9 , which is a bit array indicating which factors
are valid for a particular word.
7.3.4
Phrases
Since decoding proceeds in the translation of input phrases to output phrases, a lot of operation
involve the class Phrase10 . Since the total number of input and output factors is known to the
decoder (it has to be specified in the configuration file moses.ini), phrases are also a bit smarter
The Phrase class implements many useful functions, and two other classes are derived from it:
• The simplest form of input, a sentence as string of words, is implemented in the class
Sentence11 .
• The class TargetPhrase12 may be somewhat misleadingly named, since it not only contains a output phrase, but also a phrase translation score, future cost estimate, pointer to
source phrase, and potentially word alignment information.
7.4
Tree-Based Model Decoding
The chart decoder is a recursive variant of CKY+ parse/decoding which is able to process arbitrary context free grammars with no limitations on the number of terminals or non-terminals in
a rule. During decoding, all contiguous spans over the input spans are filled with hypotheses
(partial translations). Rules are stored in a prefix tree, which is processed incrementally, in a
way akin to Early parsing. Once rules are looked up, cube pruning is applied to pick off the
most likely applicable rules and hypotheses from underlying spans.
9
http://www.statmt.org/moses/html/df/d4d/classMoses_1_1Phrase.html
11
http://www.statmt.org/moses/html/d3/dfd/classMoses_1_1Sentence.html
12
http://www.statmt.org/moses/html/db/d67/classMoses_1_1TargetPhrase.html
10
282
7.4.1
7. Code Guide
Looping over the Spans
ChartManager::ProcessSentence13
The main loop of the decoding process fills up the stack bottom up: first looping of the width
of the span, and then over the starting position of the span.
for (size_t width = 1; width <= size; ++width) {
for (size_t startPos = 0; startPos <= size-width; ++startPos) {
For each span, first the applicable rules are created and then the rules are applied.
m_transOptColl.CreateTranslationOptionsForRange(startPos, endPos);
ChartCell &cell = m_hypoStackColl.Get(range);
cell.ProcessSentence(m_transOptColl.GetTranslationOptionList(range)
,m_hypoStackColl);
Processing a span concludes with pruning, cleaning, and sorting of the hypotheses that were
placed into the span.
cell.PruneToSize();
cell.CleanupArcList();
cell.SortHypotheses();
Consulting Rule Tables
ChartTranslationOptionCollection::CreateTranslationOptionsForRange14
Get existing data (or pointer?) from global rule collection
ChartTranslationOptionList &chartRuleColl = GetTranslationOptionList(startPos, endPos);
Multiple rule tables may be consulted. In fact, in most setups there will be a main rule table
and a rule table for glue rules. So, we need to consult each of them.
13
14
http://www.statmt.org/moses/html/d9/dde/classMoses_1_1ChartManager.html
http://www.statmt.org/moses/html/de/dae/classMoses_1_1ChartTranslationOptionCollection.html
7.4.
Tree-Based Model Decoding
283
ChartRuleLookupManager &ruleLookupManager;
ruleLookupManager.GetChartRuleCollection(wordsRange, true, chartRuleColl);
7.4.2
Looking up Applicable Rules
ChartRuleLookupManager::GetChartRuleCollection15
There are multiple implementations of the rule table. At the time of this writing, there are two:
one that is kept entirely in memory and a second one that keeps the data on disk. There is an
obvious RAM/speed trade-off here: the in-memory rule table is faster, but for some models
there may not be sufficient RAM. Both are implemented very similarly, however.
For a given span, there may be many rules that could apply. Each rule may consume input
words directly or build on any number of hypotheses that were generated for sub-spans. There
is a combinatorial explosion of sub-spans and input words that could be combined, if there is
a rule in the rule table.
The implementation of rule lookup is inspired by Early parsing, which allows for incremental
lookup. Consider the English-German rule:
PP/NP -> of the JJ_1 problem of NP_2 ; des ADJ_1 Problems NP-GEN_2
This is a good time to clarify some terminology:
•
•
•
•
•
•
PP is the source side parent non-terminal
NP is the target side parent non-terminal
JJ and NP are the source side child non-terminals
of, the, problem and of are the source side child terminals (or words)
ADJ and NP-GEN are the target side child non-terminals
des and Problems are the target side child terminals (or words)
Instead of child, the term right hand side, and correspondingly instead of parent, the term left
hand side is often used.
To check if the rule matches, we have to see if there are indeed English words of, the, problem,
and of in the input and the intervening spans have the label JJ and NP. In addition, the rule
can only apply if we have hypotheses with the constituent labels ADJ and NP-GEN in the corresponding spans.
We store the rule in a prefix structure with the following nodes:
of -> the -> JJ/ADJ -> problem -> of -> NP/NP-GEN -> des ADJ_1 Problems NP-GEN_2
15
http://www.statmt.org/moses/html/dd/d62/classMoses_1_1ChartRuleLookupManager.html
284
7. Code Guide
The key insight is the following: If there is such an applicable rule for the span under consideration, then a sub-span with the same start position matched part of this prefix path, namely:
of -> the -> JJ/ADJ -> problem -> of
If, by the time we build the prefix, we know all possible terminal or non-terminal extensions
(such as NP/NP-GEN) that match the prefix and start at the end of the sub-span, we can recursively find all rule applications, visiting each prefix and sub-span exactly once. Note: It does
not matter if there is a translation rule associated with the prefix path, all it matters that it
matches the sub-span.
We traverse the chart in a depth-first right-to-left order to ensure that all chart cells starting at
the end position of the sub-span under considerations have been processed before (and that we
can immediately find all applicable extensions of the prefix, such as NP/NP-GEN).
The code in ChartRuleLookupManagerMemory::GetChartRuleCollection16 implements the lookup
that we just described. It is extensively commented, so that is should be understandable without further explanation.
From the outside, each time the function is called for a span, it updates its internal data structures for processed rules and returns applicable rules for the span (with their target side) as a
ChartTranslationOptionList.
Let us take a closer look at the data structures used in rule lookup in the function ChartRuleLookupManagerMemo
CoveredChartSpan
Recall that processed rules are lookup up using a prefix tree, for example:
of -> the -> JJ/ADJ -> problem -> of -> NP/NP-GEN -> des ADJ_1 Problems NP-GEN_2
To use the rule, we need to know how each of these nodes in this path matches the chart, in
other words: how each node matches a span. This is encoded in a linked list of CoveredChartSpans.
Each CoveredChartSpan contains the start and end position in the span (store in a WordsRange,
which is essential a pair of integers with additional utility functions), the source word or source
span label that is matched (as a Word), and a back-pointer to the previous CoveredChartSpan
The linked list of CoveredChartSpans contains all the necessary information about applying the
rule to the source side. Target side information is stored elsewhere. Note that target-side nonterminals are not stored anymore, since they will be contained in the target side information.
The core operations of the data structure are GetSourceWord, GetWordsRange, and GetPrevCoveredChartSpan.
There is also the utility function IsNonTerminal which checks if the last word in the linked list
is a non-terminal.
16
17
http://www.statmt.org/moses/html/df/d0b/classMoses_1_1ChartRuleLookupManagerMemory.html
http://www.statmt.org/moses/html/df/d0b/classMoses_1_1ChartRuleLookupManagerMemory.html
7.4.
Tree-Based Model Decoding
285
ChartTranslationOptionList and ChartTranslationOption
A ChartTranslationOptionList is a vector of ChartTranslationOptions, with additional
utility functions.
Rules are added to this in batches, since rules are stored in the prefix tree first by matching the
source words and child non-terminals, pointing towards a list of target sides.
The list contains fully fledged out rules with target sides (opposed to the cube pruning algorithm described by Chiang, where rules are groups modulo the target side words). The list
also observes rule limits, i.e., the maximum number of rules considered for a span. It sorts
the rules by the future score of the target side (weighted translation model costs and language
model estimates), and prunes out the worst ones when the limit is exceeded. Internally a score
threshold is kept, to not even add rules that would be pruned out anyway.
be done -– the computation of the alignment between non-terminals by calling ChartTranslationOption::Create
This is done lazily, once the list finalized and pruned.
ChartTranslationOption
A ChartTranslationOption contains all the information about a rule and its application to the
span. It contains a linked list of CoveredChartSpan which details how the rule matches the
input side, a TargetPhrase with the output, and the span it applies to (in a WordsRange).
This information has to be specified at instantiation and can be queried with the functions
GetLastCoveredChartSpan, GetTargetPhrase, and GetSourceWordsRange.
Once created, the mapping between the source and target non-terminals is computed by calling
CreateNonTermIndex and can be queried with GetCoveredChartSpanTargetOrder. That information is already stored with the TargetPhrase, but it is reformated here for easier querying
(at the cost of a higher memory footprint).
PhraseDictionarySCFG
This class implements the prefix tree that contains the rules.
TargetPhraseCollection
PhraseDictionaryNodeSCFG
7.4.3
Applying the Rules: Cube Pruning
Above, we described how all the rules that apply to the current span are retrieved. In fact,
more than that is done: we also annotate each rule how it applies to the span, especially how
the non-terminals match the sub-spans.
Applying a rule now only requires the selection of the hypotheses in the specified sub-spans
that match the non-terminals in the rule.
To repeat our example:
286
7. Code Guide
PP/NP -> of the JJ_1 problem of NP_2 ; des ADJ_1 Problems NP-GEN_2
We have already identified which sub-spans the target non-terminals ADJ and NP-GEN apply to.
For each of these, however, we may have multiple choices (note that there will be at least one
for each, otherwise we would not consider the rule).
The term cube pruning derives from the fact that we explore for each rule a multi-dimensional
space, with one dimension for each non-terminal (and, in the original Chiang algorithm, but
not here, one dimension for the target phrase). This space is not always a cube, only if there are
three dimensions (two non-terminals in the Chiang algorithm), and even then it is not a cube
because the dimensions typically have differing lengths. And it is not technically a pruning
algorithm (which removes bad examples after the fact), but a greedy search for the best rule
applications. But hey, what’s in a name?
Given the cube, we sort each dimension, so that the most promising rule application is in the top
left front corner. Most promising, because for each of the non-terminals, we use the matching
hypothesis with the best score (and the target phrase with the best future cost estimate).
Finally recall that there are multiple cubes, one for each applicable rule.
ChartCell::ProcessSentence18
The cube pruning algorithm is given two data structures, a ChartTranslationOptionList that
contains all applicable rules (they are now called ChartTranslationOption) and the ChartCellCollection
that contains the chart as it has been filled in so far. The cube pruning algorithm is housed in
the ChartCell that corresponds to the span we are now filling.
First, we have to build the RuleCubeQueue. Recall, how each applicable rule has a cube. Well,
we throw them all together into one big cube (not really a regular geometrical shape, since the
dimensions differ for each rule application). Be that as it may, the first part of the algorithm
loops through the ChartTranslationOptionList and creates a RuleCube and adds it to the
cube.
The RuleCubeQueue keeps the RuleCubes sorted, so that we can always pop off the most promising rule application with its most promising underlying sub-span hypotheses. For a specified
number of times (staticData.GetCubePruningPopLimit())
•
•
•
•
we pop off the most promising RuleCube (by calling ruleCubeQueue.Pop())
build a new ChartHypothesis and calculate its score (hypo->CalcScore())
add the hypothesis to the ChartCell
ChartCell
A chart cell contains the hypothesis that were created for a span. These hypothesis are grouped
by their target side non-terminal.
18
http://www.statmt.org/moses/html/d0/dd7/classMoses_1_1ChartCell.html
7.4.
Tree-Based Model Decoding
287
RuleCubeQueue
RuleCubeQueue is a priority queue of RuleCubes (candidate rule applications). Initially it contains the top-left-front rule application for each ChartTranslationOption. When these are
expanded, their neighbors are added to the RuleCubeQueue.
Note that the same neighbor might be reached in multiple ways. If the rule applications
(0,0,0), (1,0,0) and (0,1,0) are popped off, then the latter two point to (1,1,0). This is checked
in RuleCubeQueue, which does not allow insertion of duplicates.
RuleCube
This class contains the cube for a rule. It contains information about the ChartTranslationOption
and the a list of underlying sub-span hypotheses for each non-terminal in the rule. The latter is
represented as a vector of ChildEntrys, which are essentially ordered lists of hypotheses with
When the RuleCube is created from a ChartTranslationOption, the vector of ChildEntrys is
assembled from the information in the chart. Also, the estimated score of the top-left-front rule
application is computed and stored. Note that this is a estimated score, it does not have the
real language model cost.
A RuleCube always points to a particular rule application (i.e., particular sub-span hypotheses) in the cube. If it is picked to create an hypothesis, then its neighbors are added to the
RuleCubeQueue -– this is implemented in the function CreateNeigbors. Consequently, for a
particular ChartTranslationOption, there may be multiple RuleCubes in the RuleCubeQueue.
7.4.4
Hypotheses and Pruning
New hypotheses are build in the ChartCell::ProcessSentence19 function from a RuleCube.
ChartHypothesis *hypo = new ChartHypothesis(*ruleCube, m_manager);
hypo->CalcScore();
A ChartHypothesis contains various type of information about a entry in the chart, i.e., a
translation of the covered span.
• Book-keeping
– size_t m_id hypothesis ID
– Manager& m_manager reference to manager
– WordsRange m_currSourceWordsRange covered span
– ChartTranslationOption &m_transOpt rule that created it
– vector<size_t> &m_coveredChartSpanTargetOrder covered sub-spans
• Scores
19
http://www.statmt.org/moses/html/d0/dd7/classMoses_1_1ChartCell.html
288
7. Code Guide
– ScoreComponentCollection m_scoreBreakdown all scores
– ScoreComponentCollection m_lmNGram language model scores
– ScoreComponentCollection m_lmPrefix estimated language model scores for prefix
– float m_totalScore total weighted score
• Information relevant for recombination and later use
– Phrase m_contextPrefix first words (not yet fully LM-scored)
– Phrase m_contextSuffix last words (affect later attached words)
– size_t m_numTargetTerminals length of phrase (number of words)
• Back-tracking
– ChartHypothesis *m_winningHypo points to superior hypothesis if recombined away
– ArcList *m_arcList all arcs that end at the same trellis point as this hypothesis
– vector<const ChartHypothesis*> m_prevHypos underlying hypotheses
When a hypothesis is created, the book-keeping and information relevant for recombination and later
use is set.
Scores are computed by the function CalcScore, by adding up the scores from the underlying
hypothesis, the rule application, and language model scoring of the resulting phrase. Language
model scoring (in function CalcLMScore) is a bit complex, since we do not want to re-compute
any of the language model scores that we already computed for the underlying hypotheses.
See the documented code for details.
Hypothesis recombination is handled by ChartHypothesisCollection. The function Add is
called to check if the hypothesis is recombinable with anything already in the collection (the
class ChartHypothesisRecombinationOrderer handles state matching and calls ChartHypothesis::LMContex
AddHypothesis calls Add, and handles the recombination by potentially replacing the existing
7.5
Modern server machines have multiple CPUs, each with multiple cores. Utilizing all these
cores requires either running multiple processes on the same machine or writing programs
Since many aspects of a machine translation system (training, tuning, using) lend themselves
very easily to parallel processing, Moses increasingly uses multi-threading in its components.
At this point, the following components allow for parallel execution when the switch "--threads
NUM" is added with an appropriate maximum number of threads executed at the same time:
• the decoder binary moses
• the minimum error rate training tuner mert
• the hierarchical rule extractor extract-rules
Multi-threading in Moses is based on the C++ Boost libraries, and two Moses helper libraries
that make the type of multi-threading that is typical for Moses more convenient: ThreadPool
and OutputCollector.
We will explain the implementation of multi-threaded processing on hand of a simple example.
7.5.
7.5.1
289
The part of the program that is to be run in parallel threads is called a task, and it needs to be
placed into a class of its own.
{
public:
void Run() {
std::cout << "Hello World." << endl;
}
}
Such a class now allows to be instantiated and run:
new->Run();
delete(new);
This will print "Hello World.", and is otherwise not very exciting.
Let’s make the task a bit more interesting. Our new tasks waits for a random amount of time
and then prints out a message:
void Run() {
// length of pause
int r = rand()%10;
// pause
int j = 0;
for(int i=0; i<1e8*r; i++) { j+=i; }
// write message (and length of pause)
std::cout << m_message << " (" << r << ")" << endl;
}
290
7. Code Guide
We can now create multiple instances of this task, and execute each:
srand(time(NULL));
// serial execution
This will print out three lines (the random numbers in parenthesis will vary):
zero (2)
one (4)
two (5)
Okay, where is the multi-threading? Here it comes.
7.5.2
Instead of simply running one of the tasks after the other, we assign them to a thread pool.
Once assigned, they are spawned off to a thread and will be executed in parallel to the running
main process.
// wait for all threads to finish
pool.Stop(true);
7.5.
291
That’s all too easy to be true, right? Yes, it is.
Since the three threads are running in parallel, there is no telling when they print out their
message. Not only could the lines be printed in a different order then the tasks were scheduled,
the threads may even write all over each other.
This is the catch with multi-threading: any interaction with non-local data structures must be
handled very carefully. Ideally, threads only change local data (defined in the class), and once
they are done (after pool.Stop(true)), results can be read out. This is in fact what happens in
In our case, as in the decoder, we want to output text line by line (the decoder outputs translation, and possibly additional information such as n-best lists).
7.5.3
OutputCollector
The Moses code offers the class OutputCollector to buffer up the output until it is safe to print
out. In the simplest case, it prints to STDOUT, but it can also write to a file, and indeed it offers
both regular output (default STDOUT) and debugging output (default STDERR), which both can
be redirected to different files.
Moses::OutputCollector* outputCollector = new Moses::OutputCollector();
A task can then send its output to the output collector with the function Write, for example:
m_collector->Write(id, "Hello World!");
The id is the sequential number of the sentence, starting at 0. This helps the output collector to
keep track of what can be written out and what needs to be buffered. The output collector will
not write output for sentence 1, if it has not yet received output for sentence 0.
7.5.4
By default, the Task objects are deleted after execution. However, you may want to keep the
objects around. This happens for instance in mert, where each Task finds an optimized weight
setting, which is to be processed afterwards. In this case, you have to add the following lines
virtual bool DeleteAfterExecution() {
return false;
}
20
http://www.statmt.org/moses/html/d6/d7d/mert_8cpp_source.html
292
7. Code Guide
7.5.5
Limit the Size of the Thread Queue
By default, when a thread is submitted to the ThreadPool by calling its Submit() function, it
is added to an internal queue, and the main process immediately resumes. That means, if a
which may consume a lot of memory.
If you want to restrict the number of threads in the queue, you can call, say, pool.SetQueueLimit(1000)
to limit it to 1000 queued Task instances. When the queue is full, Submit() blocks.
7.5.6
Example
Below now the complete example.
Note:
• The task class has now two more class variables which are set upon instantiation: the
sequence id m_id (a sequential number starting at 0), and a pointer to the output collector
m_collector.
• Always implement a fallback to non-threaded compilation (#ifdef WITH_THREADS .. #else
.. #endif)
• Output is placed into a file named output-file.txt (lines 43-45) instead of STDOUT.
01:
02:
03:
04:
05:
06:
07:
08:
09:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
26:
27:
28:
29:
30:
31:
#include
#include
#include
#include
#include
#include
#include
<iostream>
<fstream>
<ostream>
<cstdlib>
<sstream>
"OutputCollector.h"
using namespace std;
{
private:
unsigned int m_id;
string m_message;
Moses::OutputCollector* m_collector;
public:
ExampleTask(unsigned int id, string message, Moses::OutputCollector* collector):
m_id(id),
m_message(message),
m_collector(collector) {}
void Run() {
// length of pause
int r = rand()%10;
// pause
int j = 0;
for(int i=0; i<1e8*r; i++) { j+=i; }
7.5.
32:
33:
34:
35:
36:
37:
38:
39:
40:
41:
42:
43:
44:
45:
46:
47:
48:
49:
50:
51:
52:
53:
54:
55:
56:
57:
58:
59:
60:
61:
62:
63:
64:
65:
66:
67:
68:
69:
70:
71:
293
// write message (and length of pause)
ostringstream out;
out << m_message << " (" << r << ")" << endl;
m_collector->Write(m_id, out.str());
}
};
int main ()
{
// output into file
string outfile = "output-file.txt";
std::ofstream *outputStream = new ofstream(outfile.c_str());
Moses::OutputCollector* outputCollector = new Moses::OutputCollector(outputStream);
srand(time(NULL));
// wait for all threads to finish
pool.Stop(true);
#else
// fallback: serial execution
#endif
}
into you code directory or add paths so that they point to the moses/src directory and compile
as follows:
Make sure that the Boost libraries are in you compile paths.
When you run this example you will notice that, whatever the lengths of the pauses, the output
always appears in the correct order (i.e. zero, one, two).
294
7. Code Guide
7.6
History:
The log-linear model underlying statistical machine translation allows for the combination of
several components that each weigh in on the quality of the translation. Each component is
represented by one or more features, which are weighted, and multiplied together.
Formally, the probability of a translation e of an input sentence f is computed as
p(e|f) =
Y
hi (e, f)λi
(7.1)
i
where hi are the feature functions and λi the corresponding weights.
Note that the decoder internally uses logs, so in fact what is computed is
log p(e|f) =
X
log(hi (e, f)) λi
(7.2)
i
The tuning (Section 5.14) stage of the decoder is used to set the weights.
The following components are typically used:
•
•
•
•
•
phrase translation model (4 features, described here (Section 5.9))
language model (1 feature)
distance-based reordering model (1 feature)
word penalty (1 feature)
lexicalized reordering model (6 features, described here (Section 5.10))
One way to attempt to improve the performance of the system is to add additional feature
functions. This section explains what needs to be done to add a feature. Unless otherwise
specified the Moses code files are in the directory moses/. In the following we refer to the new
feature as xxx.
Side note: Adding a new component may imply that several scores are added. In the following,
as in the Moses source code, we refer to both components and scores as features. So, a feature
may have multiple features. Sorry about the confusion.
7.6.1
Video
There is a 10 minute video demonstrating how to create your own feature function here21 .
21
7.6.
7.6.2
295
Feature Function
The feature computes one or more values, and we need to write a feature function for it.
One important question about the new feature is, if it depends on just on the current phrase
translation, or also on prior translation decision. We call the first case stateless, the second
stateful. If the new feature is stateless, then it should inherit from the class StatelessFeatureFunction,
otherwise it should inherit from StatefulFeatureFunction
The second case causes additional complications for the dynamic programming strategy of
recombining hypotheses. If two hypotheses differ in their past translation decisions which
matters for the new feature, then they cannot be recombined.
For instance, the word penalty does only depend on the current phrase translation and is hence
stateless. The distortion features also depend on the previous phrase translation and they are
hence stateful. You can see the implementation of WordPenaltyProducer and DistortionScoreProducer
in the directory moses/FF.
However, new features are usually more complicated. For instance, it requires reading in a file
and representing it with a data structure and more complex computations. See moses/LM/SRI.h
and moses/LM/SRI.cpp for something more involved.
In the following, we assume such a more complex feature, which is implemented in its own
source files XXX.h and XXX.cpp. The feature is implemented as a class which inherits from either
StatefulFeatureFunction or StatelessFeatureFunction. So, you will write some code in
XXX.h that starts with
namespace Moses
{
class XXX : public StatefulFeatureFunction {
...
}
The class must contain the constructor:
XXX::XXX(const std::string &line)
: StatefulFeatureFunction(line)
{
....
}
The constructor must call the
StatelessFeatureFunction(...) or
StatefulFeatureFunction(...) or
296
7. Code Guide
or something that eventually calls one of this functions.
The constructor must also call the method
This is inherited from class FeatureFunction, it should NOT be overridden.
The line is the complete line from the ini file that instantiate this feature, eg.
KENLM factor=0 order=5 num-features=1 lazyken=1 path=path/file
The class must also contain the function:
This function returns true if, given a target phrase only factors in mask, the feature can be
evaluated. If the feature doesn’t need to look at words in the target phrase, always return true.
Return true if you don’t use factors.
A good example of IsUseable() is in
moses/LM/SingleFactor.cpp
This is the only necessary method the class HAS to implement. All other methods are optional.
An important function to override is
Override this function if the feature needs to load files. For example, language model classes
Many feature function needs parameters to be passed in from the ini file. For example,
7.6.
297
KENLM factor=0 order=5 num-features=1 lazyken=1 path=path/file
has the parameters factor, order, num-features, lazyken, path. To read in these parameters, override
the method
void FeatureFunction::SetParameter(const std::string& key, const std::string& value)
This method MUST call the same method in it’s parent class if the parameter is unknown, eg.
if (key == "input-factor") {
m_factorTypeSource = Scan<FactorType>(value);
} else {
StatelessFeatureFunction::SetParameter(key, value);
}
The feature function needs to be registered in Factory.cpp, FeatureRegistry():
#include "XXX.h"
...
MOSES_FNAME(XXX);
7.6.3
Stateless Feature Function
The above is all that is required to a feature function. However, it doesn’t do anything yet.
If the feature is stateless, it should override one of these methods from the class @[email protected]@ :
1. virtual void Evaluate(const Phrase &source
, const TargetPhrase &targetPhrase
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection &estimatedFutureScore) const;
2. virtual void Evaluate(const InputType &input
, const InputPath &inputPath
, ScoreComponentCollection &scoreBreakdown
, ScoreComponentCollection *estimatedFutureScore = NULL) const;
298
7. Code Guide
Or it can override one of these methods, specific to the StatelessFeatureFunction class.
3. virtual void Evaluate(const Hypothesis& hypo,
ScoreComponentCollection* accumulator) const
4. virtual void EvaluateChart(const ChartHypothesis &hypo,
ScoreComponentCollection* accumulator) const
Usually, method (1) should be overridden. See WordPenaltyProducer.cpp for a simple example using (1).
Note - Only scores evaluted in (1) is included in future cost estimation in phrase-based model
Some stateless feature functions needs to know the entire input sentence to evaluate, for example. a bag of word feature. In this case, use method (2).
Use method (3) or (4) if the feature function requires the segmentation of the source, or any
other information available from the context. Note - these methods are identical to the those
used by stateful features, except that they don’t return state.
Each stateless feature function can override 1 or more of the above methods. So far (June, 2013)
all stateless feature override only 1 method.
The methods are called at different stages in the decoding process.
* (1) is called before the search process, when the translation rule is created. This could be when the phrase-table is loaded (in the case of memory-based phrase-table
* (2) is called just before the search begins.
* (3) and (4) are called during search when hypotheses are created.
7.6.4
Stateful Feature Function
Stateful feature functions should inherit from class StatefulFeatureFunction. There are 2
class methods that can be overridden by the feature functions to score hypotheses:
5. virtual FFState* Evaluate(
const Hypothesis& cur_hypo,
const FFState* prev_state,
ScoreComponentCollection* accumulator) const = 0;
6. virtual FFState* EvaluateChart(
const ChartHypothesis& /* cur_hypo */,
int /* featureID - used to index the state in the previous hypotheses */,
ScoreComponentCollection* accumulator) const = 0;
7.6.
299
As the names suggest, (5) is used to score a hypothesis from a phrase-based model. (6) is used
to score 1 from the hierarchical/syntax model.
In addition, a stateful feature function can also override methods (1) and (2) from the base
FeatureFunction class.
For example, language models are stateful. All language model implementation should override (5) and (6). However, they should also override (1) to score the translation rule in isolation. See classes LanguageModelImplementation and LanguageModel for the implementation
of scoring language models.
Stateful feature function must also implement
const FFState* EmptyHypothesisState() const
7.6.5
Place-holder features
Some features don’t implement any Evaluate() functions. Their evaluation is more interwoven
with the creation of the translation rule, the feature function is just used as a placeholder where
Phrase-table (class PhraseDictionary), generation model (class GenerationDictionary), unknown
word feature (class UnknownWordPenaltyProducer), and input scores for confusion networks
and lattices (class InputFeature).
7.6.6
moses.ini
All feature functions are specified in the [feature] section. It should be in the format:
* Feature-name key1=value1 key2=value2 ....
For example,
KENLM factor=0 order=3 num-features=1 lazyken=0 path=file.lm.gz
Keys must be unique. There must be a key
* num-features=??
300
7. Code Guide
which specifies the number of dense scores for this feature.
The key
* name=??
is optional. If it is specified, the feature name must be unique. If it is not specified, then a name
is automatically created. All other key/value pairs are up to the feature function implementation.
7.6.7
Examples
The struck-out examples are formatted in the old Moses v.1, and before. The clear examples are
for current Moses in github.
NB. moses.ini files in the old format can still be read by the new decoder, if they just contain
the common, vanilla features (ie. no sparse features, suffix arrays, or new features that have
NB. 2 - Do NOT mix the old and new format in 1 ini file.
Phrase-tables
In-memory phrase-table (phrase-based):
PhraseDictionaryMemory num-features=5 path=phrase-table.gz input-factor=0 output-factor=0 table-limit=20
Note - The old method is relaxed about whether you add ’.gz’ to the file name; it will try it with
and without and see what exists. The new method is strict - you MUST specify ’.gz’ if the file
ends with .gz, otherwise you must NOT specify ’.gz’
Binary phrase-table (phrase-based):
PhraseDictionaryBinary num-features=5 path=phrase-table.gz input-factor=0 output-factor=0
Note - the binary phrase table consist of 5 files with the following suffixes:
7.6.
binphr.idx
binphr.srcvoc
binphr.tgtvoc
and (without word alignment):
binphr.srctree
binphr.tgtdata
or (WITH word alignment)
binphr.srctree.wa
binphr.tgtdata.wa
The path value must point to the PREFIX of the files. For example, if the files are called:
folder/pt.binphr.idx, folder/pt.binphr.srcvoc, folder/pt.binphr.tgtvoc ....
then
path=folder/pt
In-memory phrase-table (hierarchical/syntax):
PhraseDictionaryMemory num-features=5 path=phrase-table.gz input-factor=0 output-factor=0 table-limit=20
See ”In-memory phrase-table (phrase-based) for notes.
On-disk phrase-table (hierarchical/syntax):
301
302
7. Code Guide
PhraseDictionaryOnDisk num-features=5 path=phrase-table.gz input-factor=0 output-factor=0 table-limit=20
Note - the on-disk phrase-table consists of 5 files:
Misc.dat
Source.dat
TargetColl.dat
TargetInd.dat
Vocab.dat
The path value must point to the FOLDER in which these files are found.
Language models
SRILM:
SRILM factor=0 order=5 path=lm.gz
IRSTLM:
IRSTLM factor=0 order=5 path=lm.gz
KenLM:
KENLM factor=0 order=5 path=lm.gz
Lazy KenLM:
KENLM factor=0 order=5 path=lm.gz lazy=1
7.7.
303
Reordering models
LexicalReordering num-features=6 type=msd-bidirectional-fe input-factor=0 output-factor=0 path=reordering-table.msd-bidirectional-fe.0.5.0-0.gz
Misc features
New mose must have Distortion, WordPenalty, and UnknownWordPenalty explicitly in the list
of feature functions. They require no arguments, ie.
[feature]
UnknownWordPenalty
WordPenalty
Distortion
In the old moses, they were implicitly added by the decoder.
Sparse features
There are lots of ad-hoc features are currently implemented. You must look at the code and ask
the developer to see how to run them
7.7
Moses allows for sparse feature functions, i.e., feature functions that have a large, maybe unbounded, set of features, of which only a small subset applies to a given hypothesis.
ALL feature functions can contain sparse features. They don’t have to specify if or how many
sparse features they will have. Contrast this with dense features, where feature function must
specify how many scores they have at construction time.
To give an example: In addition to a regular n-gram language model, we could introduce
a discriminative bigram language model that discounts or promotes hypotheses that contain
specific bigrams. Each bigram in this feature function is its own feature with its own feature
weight.
These features cannot be tuned with MERT, but Moses has several other suitable tuning (Section 5.14) methods.
The incorporation of sparse features into the training pipeline22 is ongoing.
22
http://www.statmt.org/moses/?n=Moses.SparseFeatureTraining
304
7.7.1
7. Code Guide
Implementation
For basics, please refer to the respective section on Feature Functions (Section 7.6).
class PhraseLengthFeature : public StatelessFeatureFunction {
public:
PhraseLengthFeature(const string &line):
StatelessFeatureFunction(0, line)
{}
This creates a feature function PhraseLengthFeature with no dense features, but it can have
sparse features.
Setting feature values
As with all feature functions, sparse feature functions should implement the appropriate Evaluate()
methods described in FeatureFunctions (Section 7.6).
In the Evaluate() methods, a particular sparse score can can set using by calling the function
accumulator->PlusEquals(this, <name>, <value>);
where <name> is a key (a string) of the sparse feature.
Contrast this with setting a dense score:
accumulator->PlusEquals(this, <vector-of-values>);
7.7.2
Weights
There is no need to define a switch [weight] for the feature function. Each feature of the feature
function has its own named weight, which is a concatenation of the short name of the feature
function, and underscore (_) and its individual name for which the feature function sets.
These weights are placed into a weight file which is specified with the switch --weight-file.
For example, the target bigram feature weights (feature function short name dlmb for discriminative language model, bigrams) may have weights defined in lines such this:
7.8.
Regression Testing
305
dlmb_of:the 0.1
dlmb_in:the -0.1
dlmb_the:way 0.2
Features that do not have weights that are defined in this file are set to 0.
7.8
Regression Testing
7.8.1
Goals
The goal of regression testing is to ensure that any changes made to the decoder do not break
what has been determined to be correct, previously. The regression test suite is fast enough to
run often, but still should provide adequate confidence that nothing substantial has changed
about the internal workings of moses. The regression test suite is designed to run on most
UNIX-like systems. The regression test suite is run as part of the nightly build23 , so if you have
problems with the regression tests you should first check if the nightly build succeeded.
7.8.2
Test suite
The following regression tests are currently implemented (and many more have been added
since this list was written):
• basic-surface-only Tests basic translation, compares output strings and probabality
scores.
• basic-surface-binptable Tests binary phrase table
• consensus-decoding-surface Basic test of consensus decoding
• ptable-filtering Tests the filtering of the phrase table by estimated phrase cost, ensures that the estimated phrase cost stays the same and that the same list of phrases is
consistent. Matches pharaoh.
• multi-factor Test that moses can do translation with two factors (Currently does a very
basic test- it should be enhanced to at least include OOV words).
• multi-factor-binptable Tests factored setup with binary phrase table.
• multi-factor-drop Test of dropping words in a multi-factor model.
• nbest-multi-factor Tests n-best list generation for multi-factor models
• n-best Test n-best filtering, ensure consistency of top scores and score components. This
will require ensuring that any moses binary is capable of generating n-best lists.
• lattice-surface Tests lattice decoding
• lattice-distortion Tests lattice decoding with distortion (?)
• confusionNet-surface-only Tests confusion network decoding
23
http://www.statmt.org/moses/cruise/
306
7. Code Guide
• confusionNet-multi-factor Tests confusion network decoding with multiple factors
• lexicalized-reordering Tests lexical reordering model
• lexicalized-reordering-cn Tests lexical reordering model in combination with confusion network
• xml-markup Tests XML Markup in input to specify translations
7.8.3
Running the test suite
git clone https://github.com/moses-smt/moses-regression-tests.git
From the Moses root, run
./bjam --with-irstlm=/path/to/irst --with-cmph=/path/to/cmph --with-regtest=/path/to/moses-regression-tests -j8
This will run the regression tests in parallel (-j8) so be sure to set a number of CPUs that your
machine can handle.
If all goes well, you will see a list of the tests run, their status (hopefully pass), and a path where
the results are archived.
7.8.4
Running an individual test
You can run a specific test by providing the name followed by ".passed"
./bjam --with-irstlm=/path/to/irst --with-cmph=/path/to/cmph --with-regtest=/path/to/moses-regression-tests mert.basic.passed
The test name is the same as the directory name in /path/to/moses-regression-tests/tests .
7.8.5
How it works
The test suite invokes moses to decode a few sample phrases with well-known models. The
output from these invocations is then scraped for information (for example, the output translation of a sentence or its probability score) which is stored in a file called results.dat. These
values are then compared to a ground truth, which was established either by hand, from a prior
moses run, or from a pharaoh run.
7.8.
Regression Testing
307
This will provide a point-by-point analysis of each failure or success in the test as well as information.
Note: Since the test suite relies on the output of moses, changes to the output format may result
in broken tests. If you make changes that affect presentation only, you will need to update the
testing filters (which convert the raw moses output into the results.dat format).
7.8.6
Writing regression tests
Writing regression tests is easy, but since these tests must be able to be run anywhere, it is
important to keep a few things in mind. First, check out the regression-testing module from
the Git repository. Settle on what you would like to test in and choose a test name (henceforth,
this name will be TEST-NAME). Create a directory for it under regression testing.
Place the following into the directory regression-testing/tests/TEST-NAME:
• to-translate, which contains the text that will be translated by moses.
• moses.ini. This moses.ini file should have no absolute paths. All paths should be expressed in terms of the variables ${LM_PATH} and${MODELS_PATH}.
• The filter files, filter-stderr and filter-stdout. These files should read from STDIN
and write results of the form KEY = value to STDOUT. No other output should be generated. Numeric values (such as times) that do not require exact matches can have the
form KEY ~ value. These files are the trickiest part about writing a new regression test.
However, they allow great flexibility in verifying specific aspects of a decoding run.
• truth/results.dat This file should have the values (as produced by filter-stderr and
filter-stdout) that are expected from the test run.
If you need to add language models, phrase tables, generation tables or anything like this,
you will need to increment the required data version number in MosesRegressionTesting.pm.
Then, you will need to create a new .tgz file that contains the data for all the tests (the data
dependencies are not checked into the Git repository because they are extremely large). This
308
7. Code Guide
8
Reference
8.1
8.1.1
My system is taking a really long time to translate a sentence. What can I do
to speed it up ?
The single best thing you can do is to binarize the phrase tables and language models. See
question below also.
8.1.2
The system runs out of memory during decoding.
Filter and binarize your phrase tables. Binarize your language models using the IRSTLM. Binarize your lexicalized re-ordering table.
Binarizing the phrase table helps decrease memory usage as only phrase pairs that are needed
for each sentence are read from file into memory. Similarly for language models and lexicalized
reordering models.
This webpage (Section 4.3) tell you how to binarize the models.
8.1.3
I would like to point out a bug / contribute code.
We are always grateful for bug reports and code contribution. Send it to an existing Moses
developer you work with, or send it to Hieu Hoang at Edinburgh University.
If you want to check it code yourself, create a github account here1
•
•
•
•
•
Hieu Hoang
Nicola Bertoldi
Ondrej Bojar
Kenneth Heafield
We will probably ask to code review you a few times before giving you free reign. However,
there is less oversight if you intend to work on your own branch, rather than the trunk.
1
https://github.com/
309
310
8. Reference
8.1.4
How can I get an updated version of Moses ?
The best way is using git.
From the command line, type
git pull
Or use whatever GUI client you have.
8.1.5
What changed in the latest release of Moses?
See Releases (Section 2.4)
8.1.6
I am an undergrad/masters student looking for a project in SMT. What should
I do?
Email the mailing list with the title: ’Code monkey available. Will work for peanuts’ ! Seriously, there’s lots and lots of projects available. There has been 3-4 months projects in the past
which have made a significant contribution to the community and have been integrated into
the Moses toolkit. Your contribution will be grateful appreciated. Talk to your professor in the
first instance, then talk to us. We maintain a list of interesting projects (Section 1.3).
8.1.7
What do the 5 numbers in the phrase table mean?
See the section on phrase scoring (Section 5.9)
8.1.8
What OS does Moses run on?
It depends on which part.
The decoder can be compiled and run on Linux (32 and 64-bits), Windows, Cygwin, Mac OSX
(Intel and PowerPC). Unconfirmed reports of the decoder running on Solaris and BSD too.
The training and tuning scripts are regularly run on Linux (32 and 64-bits), and occasionally
on Mac (Intel). The whole of the Moses pipeline should also run on Windows under Cygwin,
however, this has not been confirmed. If you are able to run under Windows/Cygwin, please
let us know and we can update this FAQ.
When running on non-Linux platforms, beware of the following issues:
•
•
•
•
File system case-sensitivity
zcat, gzip command line programs missing
Old GIZA++ versions only compilable by specific gcc versions
Availability of Sun Grid Engine
Therefore, the only realistic OS to run the whole SMT pipeline on is Linux and Intel Mac.
8.1.
8.1.9
311
Can I use Moses on Windows ?
Yes. Moses compiles and runs in Cygwin exactly the same way as on Linux
There are a proviso though:
Cygwin is 32-bit, even on 64 bit windows. The binary language models (KenLM, IRSTLM)
need 64 bit to work with language models larger than about 2GB. This is the same as for 32 bit
Linux.
8.1.10
Do I need a computer cluster to run experiments?
The Moses toolkit uses SGE (Sun Grid Engine) cluster to parallelize tasks. Even though it is
not strictly necessary to use a cluster to run your experiments, it is highly advisable to get your
experiments to run faster.
The most CPU intensive task is the tuning of the weights (MERT tuning). As an indication, a
Europarl trained model, using 2000 sentences for tuning, takes 1-2 days to tune using 15 CPUs.
10-15 iterations are typical.
8.1.11
I have compiled Moses, but it segfaults when running.
Moses should not segfault, so the Moses developers would like to hear about it.
First of all, try to identify the fault yourself. The most common error is the ini file is not correct,
or the sentence input is badly formatted.
If necessary, you can debug the system by stepping through the source code. We put a lot of
effort into making the code easy to read and debug. Also, the decoder comes with Visual Studio
and XCode project file to help you debug in a GUI environment.
If you still can not find the solution, email the mailing list. Its useful to attach the ini file, the
output just before it crashes, and any other info that you think may be useful to help resolve
the problem.
8.1.12
How do I add a new feature function to the decoder?
This is now documented in its own section (Section 7.6).
8.1.13
Compiling with SRILM or IRSTLM produces errors.
Firstly, make sure SRILM/IRSTLM themselves have compiled successfully. You should see be
a libflm.a/libdstruct.a etc (for SRILM), or libirstlm.a. If these are not available, then something
went wrong. SRILM and IRSTLM are external libraries so the Moses developers have limited
say and knowledge of them.
SRI or IRST LM both have their own mailing list where you can ask questions if you have
problem compiling them. See here for details:
312
8. Reference
• SRILM2
• IRSTLM3
If Moses still does not compile successfully, look at the compile error to see where the compiler
is trying to find these external libraries. Occasionally (especially when compiling on 64-bit
machines), Moses expects the .a file in 1 sub-directory but they are in another. This is easily
solved by moving copying the .a file to the place where Moses expect it to be.
8.1.14
I am trying to use Moses to create a web page to do translation.
There is a subproject in Moses, in contrib/web , which allows you to set up a web page to
translate other web pages. Its written in Perl and the installation is non-trivial. Follow the
instructions carefully.
It doesn’t translate ad-hoc sentences. If you have some code which allow translation of ad-hoc
sentences, please share it with us !
8.1.15
How can a create a system that translate both ways, ie. X-to-Y as well as
Y-to-X ?
You need to do everything twice, and run 2 decoders. There is a lot of overlap between them,
but the toolkit is designed to go 1 way at a time.
8.1.16
PhraseScore dies with signal 11 - why?
This may happen means because you have a null byte in your data. Look at line 2 of model/lex.f2e.
Try this to find lines with null bytes in your original data:
grep -Pc ’[\000]’ <files ...>
(If your grep does not support Perl-style regular expression syntax (-P), you will have to express that a different way.)
If this turns out to be the problem, and you don’t want to run GIZA++ again from scratch, you
can try the following:
First go into working-dir/model and delete everything but the following:
2
3
http://www.speech.sri.com/mailman/listinfo/srilm-user
https://list.fbk.eu/sympa/subscribe/user-irstlm
8.1.
313
aligned.grow-diag-final-and
aligned.0.fr
aligned.0.en
lex.0-0.n2f
lex.0-0.f2n
Now run this fragment of Perl:
perl -i.BAD -pe ’s/[\000]/NULLBYTE/g;’ aligned.0* lex.0*
This will replace every null byte in those four files, saving the old version out to *.BAD. (This
may be overkill, for instance if only the foreign side has the problem.
Now restart the Moses training script with the same invocation as before, but tell it to start at
step 5:
train-model.perl ... --first-step 5
8.1.17
Does Moses do Hierarchical decoding, like Hiero etc?
Yes. Check the Syntax Tutorial (Section 3.3).
8.1.18
Can I use Moses in proprietary software ?
Basically, if you are just using Moses unchanged, there are no license issues. You can also
use the Moses library (libmoses.a) in your own applications. But if you want to distribute a
modified version of Moses, you have to distribute the source code to the modifications.
8.1.19
GIZA++ crashes with error "parameter ’coocurrencefile’ does not exist."
You have a version of GIZA++ which does not support cooccurrence files. To add support for
cooccurrence files, you need to edit the GIZA++ Makefile and add the flag -DBINARY_SEARCH_FOR_TTABLE
to CFLAGS_OPT. Then you should rebuild GIZA++.
4
314
8.1.20
8. Reference
Running regenerate-makefiles.sh gives me lots of errors about *GREP and
*SED macros
You should not be running this script. Moses moved from autotools to bjam in Autumn 2011.
8.1.21
Running training I got the following error "*** buffer overflow detected ***:
../giza-pp/GIZA++-v2/GIZA++ terminated"
This error occurs during the word alignment step and is related to GIZA++, and not directly to
the Moses Toolkit. Neverthless, the solution is described here5 .
8.1.22
I retrained my model and got different BLEU scores. Why?
In general, Machine Translation training is non-convex. this means that there are multiple
solutions and each time you run a full training job, you will get different results. In particular,
you will see different results when running GIZA++ (any flavour) and MERT.
The best way to deal with this (and most expensive) would be to run the full pipe-line, from
scratch and multiple times. This will give you a feel for variance --differences in results. In
general, variance arising from GIZA++ is less damaging than variance from MERT.
To reduce variance it is best to use as much data as possible at each stage. It is possible to
reduce this variability by using better machine learning, but in general it will always be there.
Another strategy is to fix everything once you have a set of good weights and never rerun
MERT. Should you need to change say the language model, you will then manually alter the
associated weight. This will mean stability, but at the obvious cost of generality. it is also ugly.
See Clark et al.6 for a discussion of some of these issues.
8.1.23
I specified ranges for mert weights, but it returned weights which are outwith those ranges
The ranges that you pass to mert-moses.pl (using the --range argument) are only used in the
random restarts, so serve to guide mert rather than restrict it.
8.1.24
Who do I ask if my question has not been answered by this FAQ?
Search the mailing list archive7 . If you still do not find the answer, then send questions to the
5
http://www.statmt.org/moses/?n=Moses.GizappBufferOverflow
http://www.cs.cmu.edu/~jhclark/pubs/significance.pdf
7
http://blog.gmane.org/gmane.comp.nlp.moses.user
8
http://mailman.mit.edu/mailman/listinfo/moses-support
6
8.2.
Reference: All Decoder Parameters
8.2
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
315
Reference: All Decoder Parameters
-beam-threshold (b): threshold for threshold pruning
-cache-path: ?
-config (-f): location of the configuration file
-constraint: Target sentence to produce
-cube-pruning-diversity (-cbd): How many hypotheses should be created for each
coverage. (default = 0)
-cube-pruning-pop-limit (-cbp): How many hypotheses should be popped for each
stack. (default = 1000)
-distortion: configurations for each factorized/lexicalized reordering model.
-distortion-file: source factors (0 if table independent of source), target factors, location of the factorized/lexicalized reordering tables
-distortion-limit (-dl): distortion (reordering) limit in maximum number of words (0
= monotone, -1 = unlimited)
-drop-unknown (-du): drop unknown words instead of copying them
-early-discarding-threshold ([email protected]@): threshold for constructing hypotheses based
on estimate cost
-factor-delimiter (-fd): specify a different factor delimiter than the default
-generation-file: location and properties of the generation table
-include-alignment-in-n-best: include word alignment in the n-best list. default is
false
-input-factors: list of factors in the input
-input-file (-i): location of the input file to be translated
-inputtype: text (0), confusion network (1) or word lattice (2)
-labeled-n-best-list: print out labels for each weight type in n-best list. default is true
-lmodel-dub: dictionary upper bounds of language models
-lmodel-file: location and properties of the language models
-lmstats (-L): (1/0) compute LM backoff statistics for each translation hypothesis
-mapping: description of decoding steps
-max-partial-trans-opt: maximum number of partial translation options per input
span (during mapping steps)
-max-phrase-length: maximum phrase length (default 20)
-max-trans-opt-per-coverage: maximum number of translation options per input span
(after applying mapping steps)
-mbr-scale: scaling factor to convert log linear score probability in MBR decoding (default 1.0)
-mbr-size: number of translation candidates considered in MBR decoding (default 200)
-minimum-bayes-risk (-mbr): use miminum Bayes risk to determine best translation
-monotone-at-punctuation (-mp): do not reorder over punctuation
-n-best-factor: factor to compute the maximum number of contenders (=factor*nbestsize). value 0 means infinity, i.e. no threshold. default is 0
-n-best-list: file and size of n-best-list to be generated; specify - as the file in order to
write to STDOUT
-output-factors: list if factors in the output
-output-search-graph (-osg): Output connected hypotheses of search into specified
filename
316
8. Reference
• -output-word-graph (-owg): Output stack info as word graph. Takes filename, 0=only
hypos in stack, 1=stack + nbest hypos
• -persistent-cache-size: maximum size of cache for translation options (default 10,000
input phrases)
• -phrase-drop-allowed (-da): if present, allow dropping of source words
• -print-alignment-info: Output word-to-word alignment into the log file. Word-toword alignments are taken from the phrase table if any. Default is false
• -print-alignment-info-in-n-best: Include word-to-word alignment in the n-best list.
Word-to-word alignments are takne from the phrase table if any. Default is false
• -recover-input-path (-r): (confusion net/word lattice only) - recover input path corresponding to the best translation
• -report-all-factors: report all factors in output, not just first
• -report-segmentation (-t): report phrase segmentation in the output
• -search-algorithm: Which search algorithm to use. 0=normal stack, 1=cube pruning
(default = 0)
• -stack (-s): maximum stack size for histogram pruning
• -stack-diversity (-sd): minimum number of hypothesis of each coverage in stack (default 0)
• -time-out: seconds after which is interrupted (-1=no time-out, default is -1)
• -translation-details (-T): for each best translation hypothesis, print out details about
what source spans were used, dropped
• -translation-option-threshold (-tot): threshold for translation options relative to
best for input phrase
• -ttable-file: location and properties of the translation tables
• -use-alignment-info: Use word-to-word alignment: actually it is only used to output
the word-to-word alignment. Word-to-word alignments are taken from the phrase table
if any. Default is false.
• -use-persistent-cache: cache translation options across sentences (default true)
• -verbose (-v): verbosity level of the logging
• -weight-d (-d): weight(s) for distortion (reordering components)
• -weight-e (-e): weight for word deletion
• -weight-file (-wf): file containing labeled weights
• -weight-generation (-g): weight(s) for generation components
• -weight-i (-I): weight for word insertion
• -weight-l (-lm): weight(s) for language models
• -weight-t (-tm): weights for translation model components
• -weight-w (-w): weight for word penalty
• -xml-input (-xi): allows markup of input with desired translations and probabilities.
values can be ’pass-through’ (default), ’inclusive’, ’exclusive’, ’ignore’
8.3
Reference: All Training Parameters
• --root-dir -- root directory, where output files are stored
• --corpus -- corpus file name (full pathname), excluding extension
• --e -- extension of the English corpus file
8.3.
Reference: All Training Parameters
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
•
8.3.1
317
--f -- extension of the foreign corpus file
--lm -- language model: <factor>:<order>:<filename> (option can be repeated)
--first-step -- first step in the training process (default 1)
--last-step -- last step in the training process (default 7)
--parts -- break up corpus in smaller parts before GIZA++ training
--corpus-dir -- corpus directory (default $ROOT/corpus) --lexical-dir -- lexical translation probability directory (default$ROOT/model)
--model-dir -- model directory (default $ROOT/model) --extract-file -- extraction file (default$ROOT/model/extract)
--giza-f2e -- GIZA++ directory (default $ROOT/giza.$F-$E) --giza-e2f -- inverse GIZA++ directory (default$ROOT/giza.$E-$F)
--alignment -- heuristic used for word alignment: intersect, union, grow, grow-final,
grow-diag, grow-diag-final (default), grow-diag-final-and, srctotgt, tgttosrc
--max-phrase-length -- maximum length of phrases entered into phrase table (default
7)
--giza-option -- additional options for GIZA++ training
--verbose -- prints additional word alignment information
--no-lexical-weighting -- only use conditional probabilities for the phrase table, not
lexical weighting
--parts -- prepare data for GIZA++ by running snt2cooc in parts
--direction -- run training step 2 only in direction 1 or 2 (for parallelization)
--reordering -- specifies which reordering models to train using a comma-separated
list of config-strings, see FactoredTraining.BuildReorderingModel (Section 5.10). (default
distance)
--reordering-smooth -- specifies the smoothing constant to be used for training lexicalized reordering models. If the letter "u" follows the constant, smoothing is based on
actual counts. (default 0.5)
--alignment-factors ---translation-factors ---reordering-factors ---generation-factors ---decoding-steps --
Basic Options
A number of parameters are required to point the training script to the correct training data. We
will describe them in this section. Other options allow for partial training runs and alternative
settings.
As mentioned before, you want to create a special directory for training. The path to that
directory has to be specified with the parameter --root-dir.
The root directory has to contain a sub directory (called corpus) that contains the training data.
The training data is a parallel corpus, stored in two files, one for the English sentences, one for
the foreign sentences. The corpus has to be sentence-aligned, meaning that the 1624th line in
the English file is the translation of the 1624th line in the foreign file.
Typically, the data is lowercased, no empty lines are allowed, and having multiple spaces between words may cause problems. Also, sentence length is limited to 100 words per sentence.
318
8. Reference
The sentence length ratio for a sentence pair can be at most 9 (i.e, having a 10-word sentence
aligned to a 1-word sentence is disallowed). These restrictions on sentence length are caused
by GIZA++ and may be changed (see below).
The two corpus files have a common file stem (say, euro) and extensions indicating the language (say, en and de). The file stem (--corpus-file), and the language extensions (--e and
--f) have to be specified to the training script.
In summary, the training script may be invoked as follows:
train-model.perl --root-dir . --f de --e en --corpus corpus/euro >& LOG
After training, typically the following files can be found in the root directory (note the time
stamps that tell you something about how much time was spent on each step took for this
data):
> ls -lh *
-rw-rw-r--
1 koehn
user
110K Jul 13 21:49 LOG
corpus:
total 399M
-rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--
1
1
1
1
1
1
1
1
1
1
koehn
koehn
koehn
koehn
koehn
koehn
koehn
koehn
koehn
koehn
user
user
user
user
user
user
user
user
user
user
104M
4.2M
3.2M
2.6M
104M
1.1M
793K
614K
94M
84M
giza.de-en:
total 422M
-rw-rw-r--rw-rw-r--rw-rw-r--
1 koehn
1 koehn
1 koehn
user
user
user
107M Jul 13 03:57 de-en.A3.final.gz
314M Jul 12 20:11 de-en.cooc
2.0K Jul 12 20:11 de-en.gizacfg
giza.en-de:
total 421M
-rw-rw-r--rw-rw-r--rw-rw-r--
1 koehn
1 koehn
1 koehn
user
user
user
107M Jul 13 11:03 en-de.A3.final.gz
313M Jul 13 04:07 en-de.cooc
2.0K Jul 13 04:07 en-de.gizacfg
model:
Jul
Jul
Jul
Jul
Jul
Jul
Jul
Jul
Jul
Jul
12
12
12
12
12
12
12
12
12
12
19:58
19:56
19:42
19:42
19:59
19:56
19:56
19:56
18:08
18:08
de-en-int-train.snt
de.vcb
de.vcb.classes
de.vcb.classes.cats
en-de-int-train.snt
en.vcb
en.vcb.classes
en.vcb.classes.cats
euro.de
euro.en
8.3.
Reference: All Training Parameters
total 2.1G
-rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--rw-rw-r--
1
1
1
1
1
1
1
1
1
koehn
koehn
koehn
koehn
koehn
koehn
koehn
koehn
koehn
user
user
user
user
user
user
user
user
user
319
94M
84M
90M
214M
212M
78M
78M
862
1.2G
Jul
Jul
Jul
Jul
Jul
Jul
Jul
Jul
Jul
13
13
13
13
13
13
13
13
13
19:59
19:59
19:59
20:33
20:35
20:23
20:23
21:49
21:49
aligned.de
aligned.en
aligned.grow-diag-final
extract.gz
extract.inv.gz
lex.f2n
lex.n2f
pharaoh.ini
phrase-table
Summary
•
•
•
•
•
8.3.2
--root-dir -- root directory, where output files are stored
--corpus -- corpus, expected in ROOT/corpus --e -- extension of the English corpus file --f -- extension of the foreign corpus file --lm -- language model file Factored Translation Model Settings More on factored translation models in the Overviev (Section 5.1). Summary • • • • • 8.3.3 --alignment-factors ---translation-factors ---reordering-factors ---generation-factors ---decoding-steps -- Lexicalized Reordering Model More on lexicalized reording on the description of Training step 7: build reordering model (Section 5.10). Summary • --reordering -• --reordering-smooth -- 8.3.4 Partial Training You may have better ideas how to do word alignment, extract phrases or score phrases. Since the training is modular, you can start training at any of the seven training steps --first-step and end it at any subsequent step --last-step. Again, the nine training steps are: 320 8. Reference 1. 2. 3. 4. 5. 6. 7. 8. 9. Prepare data Run GIZA++ Align words Get lexical translation table Extract phrases Score phrases Build reordering model Build generation models Create configuration file For instance, if you may have your own method to generate a word alignment, you want to skip these training steps and start with lexical translation table generation, you may specify this by train-model.perl [...] --first-step 4 Summary • --first-step -- first step in the training process (default 1) • --last-step -- last step in the training process (default 7) 8.3.5 File Locations A number of parameters allow you to break out of the rigid file name conventions of the training script. A typical use for this is that you want to try alternative training runs, but there is no need to repeat all the training steps. For instance, you may want to try an alternative alignment heuristic. There is no need to rerun GIZA++. You could copy the necessary files from the corpus and the giza.* directories into a new root directory, but this takes up a lot of additional disk space and makes the file organization unnecessarily complicated. Since you only need a new model directory, you can specify this with the parameter --model-dir, and stay within the precious root directory structure: train-model.perl [...] --first-step 3 --alignment union --model-dir model-union The other parameters for file and directory names fullfill similar purposes. Summary • • • • --corpus-dir -- corpus directory (defaultROOT/corpus)
--lexical-dir -- lexical translation probability directory (default $ROOT/model) --model-dir -- model directory (default$ROOT/model)
--extract-file -- extraction file (default $ROOT/model/extract) 8.3. Reference: All Training Parameters 321 • --giza-f2e -- GIZA++ directory (default$ROOT/giza.
F−
E}
• --giza-e2f -- inverse GIZA++ directory (default \$ROOT/giza.
E−
F)
8.3.6
Alignment Heuristic
A number of different word alignment heuristics are implemented, and can be specified with
the parameter --alignment. The options are:
• intersect -- the intersection of the two GIZA++ alignments is taken. This usually creates
a lot of extracted phrases, since the unaligned words create a lot of freedom to align
phrases.
• union -- the union of the two GIZA++ alignments is taken
• grow-diag-final -- the default heuristic
• grow-diag -- same as above, but without a call to function FINAL() (see background to
word alignment).
• grow -- same as above, but with a different definition of neighboring. Now diagonally
• grow -- no diagonal neighbors, but with FINAL()
Different heuristic may show better performance for a specific language pair or corpus, so some
experimentation may be useful.
Summary
• --alignment -- heuristic used for word alignment: intersect, union, grow, grow-final,
grow-diag, grow-diag-final (default)
8.3.7
Maximum Phrase Length
The maximum length of phrases is limited to 7 words. The maximum phrase length impacts
the size of the phrase translation table, so shorter limits may be desirable, if phrase table size
is an issue. Previous experiments have shown that performance increases only slightly when
including phrases of more that 3 words.
Summary
• --max-phrase-length -- maximum length of phrases entered into phrase table (default
7)
8.3.8
GIZA++ Options
GIZA++ takes a lot of parameters to specify the behavior of the training process and limits on
sentence length, etc. Please refer to the corresponding documentation for details on this.
Parameters can be passed on to GIZA++ with the switch --giza-option.
322
8. Reference
For instance, if you want to the change the number of iterations for the different IBM models
to 4 iterations of Model 1, 0 iterations of Model 2, 4 iterations of the HMM Model, 0 iterations
of Model 3, and 3 iterations of Model 4, you can specify this by
train-model.perl [...] --giza-option m1=4,m2=0,mh=4,m3=0,m4=3
Summary
• --giza-option -- additional options for GIZA++ training
8.3.9
Dealing with large training corpora
Training on large training corpora may become a problem for the GIZA++ word alignment
tool. Since it stores the word translation table in memory, the size of this table may become
too large for the available RAM of the machine. For instance, the data sets for the NIST ArabicEnglish and Chinese-English competitions require more than 4 GB of RAM, which is a problem
for current 32-bit machines.
This problem can be remedied to some degree by a more efficient data structure in GIZA++,
which requires the run of snt2cooc in advance on the corpus in parts and the merging on
the resulting output. All you need to know is that running the training script with the option
--parts n, e.g. --parts 3 may allow you to train on a corpus that was too large for a regular
run.
Somewhat related to this problem caused by large training corpora is the problem of the large
run time of GIZA++. It is possible to run the two GIZA++ separately on two machines with
the switch --direction. When running one of the runs on one machine with --direction 1
and the other run on a different machine or CPU with --direction 2, the processing time for
training step 2 can be cut in half.
Summary
• --parts -- prepare data for GIZA++ by running snt2cooc in parts
• --direction -- run training step 2 only in direction 1 or 2 (for parallelization)
8.4
Glossary
(based on excerpts from the "DoMY Glossary" in Do Moses Yourself Community Edition by Precision
Translation Tools Co., Ltd.)
This glossary includes common terms that are helpful for new users of statistical machine translation (SMT) and the open source Moses Decoder project.
8.4.
Glossary
323
aligned data: Aligned data are the elements of a parallel corpus consisting of two or more
languages. Each element in one language matches the corresponding element in the other language(s). The elements, sometimes called segments, can be block-aligned, paragraph-aligned,
sentence-aligned, phrase-aligned or token-aligned.
alignment process: There are two alignment processes. In corpus preparation, the alignment
process creates aligned data. During training, the alignment process uses a program such as
MGIZA++ to create word alignment files.
BLEU score: BLEU stands for Bi-Lingual Evaluation Understudy". A BLEU score indicates how
closely the token sequences in one set of data, such as machine translation output, correlate
with (match) the token sequences in another set of data, such as a reference human translation.
See: evaluation process
corpus preparation: Corpus preparation is the general process to extract, transform, categorize
various documents from their original purpose to and align the resulting data into a parallel
corpus for training a translation model.
development (dev) set: See "tuning set"
eval set: See "test set"
evaluation process: The evaluation process uses a translation model of components created
in the training process and configured with the tuning process to translate several thousand
source language sentences in the eval set. This process then compares the resulting machine
translations to reference translations, also in the eval set. The final BLEU score evaluation
report shows how well the machine translations match the reference translations.
hierarchical model: SMT translation model that uses hierarchical training corpus.
hierarchical training data: A training corpus with each phrase annotated with the hierarchical
structure of the language, such as parts of speech, word function, etc.
language model: A "language model" or "lm" is a statistical description of one language that
includes the frequencies of token-based n-grams occurrences in a corpus. The "lm" is trained
from a large monolingual corpus and saved as a file. The language model file is a required
component of every translation model. Moses uses language model to select the most "probably" target language sentence from a large set of "possible" translations it generated using the
phrase table and reordering table.
language model types: Language model files contain statistical data generated by one of various programs. Moses Decoder can use language model file types including: KenLM SRILM,
RandLM and IRSTLM. SRILM, RandLM and IRSTLM toolkits include tools that train the new
language model files. KenLM, however, only reads ARPA standard language model files which
can be created by SRILM, IRSTLM.
Moses configuration file: The Moses configuration file is a text file created during the tuning
process. The file contains the paths to the phrase table(s), reordering table, language model(s)
with other codes and numeric values that control how the Moses Decoder works.
n-grams: An n-gram is a subsequence of n number of (1, 2, 3, etc) items in a larger sequence.
In an lm n-grams are sequences of tokens. In phrase tables and reordering tables, n-grams are
sequences of pairs of source and target language tokens.
parallel corpus or parallel data: A linguistic corpus of two or more languages where each
element in one language corresponds to an element with the same meaning in the other language(s). The original, authored language is identified as the source language. Non-source
324
8. Reference
languages are referred to as "target" languages. For Moses SMT, parallel data takes the form of
one source and one target language text file where both files contain corresponding translation
of sentences line by line.
phrase table: A "phrase table" is a statistical description of a parallel corpus of source-target
language sentence pairs. The frequencies that n-grams in a source language text co-occur with
n-grams in a parallel target language text represent the probability that those source-target
paired n-grams will occur again in other texts similar to the parallel corpus. In practical terms,
the phrase table is a file created during the training process and saved in the translation model
folder. It functions as a sophisticated dictionary between the source and target languages.
Phrase tables and reordering tables are translation model components.
pipeline: A "pipeline" is a toolchain of processes connected by standard streams, so that the
output of each process (stdout) feeds directly as input (stdin) to the next one.
recaser model: A recaser model is a special translation model translates lower cased data to
"natural" cased text (upper and lower casing).
reordering table: A "reordering table" contains the statistical frequencies that describe the
changes in word order between source and target languages, such as "big house" versus "house
big". In practical terms, a "reordering table" is a file created during the training process and
saved as a file in the model folder. The reordering table is translation model components.
source language: The source language is the language of the text that is to be translated. Typically, this is the authored language of the text. The source language is the same as the TMX
specification "srclang" attribute of the <tu> tag.
target language: The target language is the language the source language text should be translated to.
test set: A pair of source and target language data, typically containing of several thousands of
pairs used in the evaluation process.
tokenization: Tokenization is the process of separating words from punctuation and symbols
into tokens.
tokens: Tokens are the basic unit in a machine translation process. Tokens are a sequence of
characters, such as words, punctuation or symbols, separated by a space. See: BLEU score
toolchain: A "toolchain" is a series of linked or "chained" programming tools used in a series
where the output of an upstream tool become the input for a "downstream" tool.
training corpus or training data: A linguistic corpus with parallel data prepared for training
into the phrase table and a reordering table components of a translation model.
training process: Training is a process in the machine learning branch of artificial intelligence
field. In the training process, a system "learns" the relationships between parallel data. In SMT,
the source language texts are stimuli that generate the target language text as a response. In
practical terms, training starts with the bitext files and creates the phrase table and reordering
table that are components of a translation model.
translation memory: A translation memory (TM) is parallel data that was collected for the
purpose of aiding future translations.
translation model: A "translation model" consists of one or more phrase tables, zero or more
reordering tables, one or more language models and one Moses configuration file that were
created during the training and tuning processes.
8.4.
Glossary
325
truecasing: Truecasing replaces each words with its natural uppercase/lowercase spelling.
This process typically leaves all words unchanged except for the first word in the sentence,
which may be lowercased.
tuning process: Tuning is a process that finds the optimized configuration file settings for
a translation model when used a specific purpose. The tuning process translates thousands
of source language phrases in the tuning set with a translation model, compares the model’s
output to a set of reference human translations, and adjusts the settings with the intention to
improve the translation quality. This process continues through numerous iterations. With
each iteration, the tuning process repeats the steps until it reaches an optimized translation
quality.
tuning set: A pair of source and target language data, typically containing of several thousands
of pairs used in the tuning process.
word aligner: A word aligner is a program that created word alignment files during the word
alignment process. Moses currently supports these word aligners: GIZA++, MGIZA++, and
BerkeleyAligner.
word alignment: Word alignment process uses a word aligner to create a word alignment file
during the training process.
words: A word is the smallest unit of meaning in a language that will stand on its own. In
SMT, a word is a token created in the tokenization process that is not a punctuation or symbol.