Text Beautifier: An Affective-Text Tool to Tailor Written Text

Document technical information

Format pdf
Size 263.8 kB
First found May 22, 2018

Document content analysis

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

Persons

Organizations

Places

Transcript

Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (2008)
Text Beautifier: An Affective-Text Tool to Tailor Written Text
Fahim Kawsar1, Mostafa Al Masum Shaikh2, Mitsuru Ishizuka2
1
2
Department of Computer Science, Waseda University, Tokyo, Japan
Department of Information and Communication Engineering, University of Tokyo, Tokyo, Japan
[email protected], [email protected]
niques to score a new word utilizing commonsense knowledge and online thesaurus. Grounding on our textual affect
sensing approach we can obtain a computational model of
an input sentence. Hence the model can be exploited with
the list of scored words to develop a tool that will help a
user to tailor an input sentence in different magnitudes. For
example using such a tool one may get the recommendations like, “Your lovely email makes me very glad”, or “I
become glad to read your email”, or “I am very happy to
obtain your nice email” for the input sentence, “I am happy
to receive your email” after scaling up the input sentence
on affect, or prospect, or valence scale respectively.
Abstract
We have spelling and grammar checking tools available on
today’s word processors. But what they are missing is a tool
that can recommend several possibilities of a given written
sentence to assist a user to write better sentences. Therefore, we aim to develop a linguistic tool to beautify text by
applying our developed lexical resources regarding textual
affect sensing. The developed tool will allow a user to beautify an input sentence in terms of tuning it on different
scales like valence, affect, prospect, and praise. For example
using such a tool one may get the recommendations like,
“Your lovely email makes me very glad”, or “I become glad
to read your email”, or “I am very happy to obtain your nice
email” for the input sentence “I am happy to receive your
email” after scaling up the input sentence on affective, or
prospective, or valence scale respectively. Such tool will be
especially helpful to the non-native English speakers to
write better English.
The Core Idea
Let’s assume that someone received an email from a journal editor regarding one’s submission and one is preparing
to answer that email like this, “Thank you for your email. I
am going to revise the paper according to the feedbacks
given by the reviewers. I plan to send the revised paper in a
week time”. In order to explain the core idea of this paper
we take this example text, which the email responder plans
to beautify. Our approach is explained as following.
Computational Model: We have created a Semantic
Parser that provides the computational model of the input
sentence by outputting triplet(s). A triplet consists of a subject or agent, a verb, and an object. Each member of the
triplet may or may not have associated attribute(s) (e.g. adjective, adverb etc.). We first obtain XML-formatted
syntactic and functional dependency information of each
word of the input sentence using the Machinese Syntax
parser and this output constitutes the basis for further
processing to generate the triplet(s). Since a triplet is
initiated with an occurrence of a verb in the sentence, the
semantic parsing may obtain more than one such triplet if
the there are multiple verbs in the sentence. Basically a
triplet encodes information about “who is associated with
what and how” with a notion of semantic verb frame
analysis. For example, the input sentence “I am going to
revise the paper according to the feedbacks given by the
reviewers.”, produces three triplets as shown in Table 1.
Introduction
Analyzing the relationship between the components (e.g.,
words, clause, etc.) of a written text and the affect conveyed by the underpinned meaning of the text, a program
can recommend several representations of the given text.
We have spelling and grammar checking tools available on
today’s word processors to write grammatically correct
sentences. But what they are missing is a tool that can recommend several possibilities of a given written sentence to
assist a user to write better sentences. Therefore, we aim to
develop a linguistic tool, Text-Beautifier, which will assist
a user to beautify an input sentence in terms of tuning it on
different scales like valence, affect, prospect, and praise.
The idea of such scales is adopted from the OCC (Ortony,
Clore, and Collins 1988) theory of emotion. Our developed
system1 can sense the possible emotions being conveyed
by an input sentence. For example, for the input, “I am
looking forward to your answer”, our system outputs that
the given sentence carries positive valenced emotions like,
‘Joy’, ‘Hope’, and ‘Gratitude’. The output is achieved by a
rule-based approach of computing contextual-valence values to the linguistic components of the input sentence.
While developing this affect-sensing program we have created lexical resources comprising of a growing list of valence assigned words. Moreover we have devised tech1
Table 1: Triplets of a Sample Sentence
[['Subject Name:', 'I', 'Subject Type:', 'Self', 'Sub Attrib:',
[]], ['Action Name:', 'going to', 'Action Status:', 'Present
Continuous ', 'Action Attrib:', ['dependency: to']], ['Object
http://www.almasum.com/research/emotest
1806
crease the strength of the sentence our approach is to reduce or increase the contextual valence of the Triplets (Table 1) by picking up lower or upper valenced synonyms.
Affect Scale: The strength of the affective scale is calculated according to the output of our system [2] which gives
the named emotional affinity of the input sentence. The example sentence shows the affinity with “Joy (57.5%)”,
“Happy-for (46%)”, “Hope (34.5%)”, “Love (23%)” and
“Satisfaction (11.5%)” emotions according to our system.
Since the maximum strength is found to be 57.5%, the sentence gets the value of 6 on affective scale. In order to
make the sentence stronger on affective scale our approach
is to add appropriate and higher valenced adjectives with
the object (e.g., prudent reviewer) of a triplet. Similarly, to
make the sentence weaker on affective scale our approach
is to either eliminate the adjectives or use lower valenced
adjectives without reversing the global sentiment of the
sentence.
Prospect Scale: “Prospect” and “Praise” values indicate
the lexical affinity of an action word with respect to “desirable or undesirable”, and “praiseworthiness or blameworthiness”, respectively. According to the OCC model, prospect of an event involves a conscious expectation that it
will occur in the future, and the value can be either “positive” or “negative”. For example, for the events like, “admit to university”, “revise paper”, and “give feedback”, we
get +9.375, +8.728, and +4.921 as the valence values for
the events, respectively. So in order to make a sentence
more prospective our target is to increase the prospective
value of the events represented by the triplets.
Praise Scale: According to the appraisal structure of an
event mentioned in the OCC model, an event can be assessed as either “praiseworthy” or “blameworthy”. Therefore we have either a positive or negative score for a verb
to associate “praise” or “blame” to the list of action words
without a context. For example, let’s consider these events,
“pass final exam”, “forget friend’s birthday”, and “revise
paper”. For these events we get +7.95, -9.31, and +3.87,
respectively as the scores of the events on praise scale.
Therefore our approach to tailor the sentence on this scale
is to select the appropriately valenced synonymous verbs
without revering the meaning of the sentence.
Name:', '', 'Object Type:', '', 'Object Attrib:', []]]
[['Subject Name:', 'I', 'Subject Type:', '', 'Sub Attrib:', []],
['Action Name:', 'revise', 'Action Status:', 'Infinitive ', 'Action Attrib:', ['dependency: by']], ['Object Name:', 'paper',
'Object Type:', 'N NOM SG', 'Object Attrib:', ['Determiner: the', 'PREP: according to']]]
[['Subject Name:', 'reviewer', 'Subject Type:', 'N NOM
SG', 'Sub Attrib:', []], ['Action Name:', 'give', 'Action
Status:', 'passive Past Particle', 'Action Attrib:', []], ['Object Name:', 'feedback', 'Object Type:', 'N NOM PL ', 'Object Attrib:', ['Determiner: the']]]
GUI: The GUI of the prototype application is shown in
Figure 1. The system takes a sentence as an input. Clicking
on “Measure” button does the affect sensing according to
our system [2] and measure the valence, affect, prospect
and praise values of the input sentence to be shown on the
“Beauty-Scale” panel represented by four scales. A user
can interact with the sliders of the scales by dragging to
right or left. For example, if a user wants to make the input
sentence more affective but more valenced, he drags the
sliders of the “Affect Scale” and “Valence Scale” to right
respectively. Each scale varies from 1 to 10, which represent the overall strength of the sentence on that particular
scale. How the strength is calculated is explained in the
next subsections. By moving the sliders back and forth a
user can observer the recommended sentence to accept or
tailor further.
Conclusion
Figure 1: Prototype Application GUI
At present we are revising the valence values of the words
and developing the rules to incorporate different subtle features of writing well structured sentences. Such tool will be
especially helpful to the non-native English speakers to
write better English.
Valence Scale: The valence scale’s strength is calculated
according to our developed tool SenseNet, which calculates the contextual-valence of the words using rules and
prior-valence values of the words of the input sentence. It
outputs either a negative or positive value that indicates a
numerical measure of negative and positive sentiments carried by the sentence. For the ongoing example input we get
+7.44 as the strength on the valence scale. This indicates
that the sentence is a positive one and the beautification
will occur towards the positive direction while interacting
with the sliders on each scale. In order to increase or de-
References
Ortony, A., Clore, G.L. and Collins, A. 1988. The Cognitive
Structure of Emotions, Cambridge University Press
1807

Similar documents

×

Report this document