Previous Researches on Lexical Ambiguity and Polysemy

Document technical information

Format pdf
Size 125.5 kB
First found May 22, 2018

Document content analysis

Category Also themed
not defined
no text concepts found





Chapter 2
Previous Researches on Lexical Ambiguity
and Polysemy
The goal of this study aims to explore all possible senses of the four target words,
which are all lexically ambiguous words. In this chapter, I would like to introduce
and discuss lexical ambiguity and polysemy. In addition, I also would like to
discuss that lexical ambiguity studies are used in the corpus-based and computational and psycholinguistic approaches. Therefore, I will discuss corpus-based and
computational models and experimental evaluation of the psycholinguistic perspective for the four target words in this sense prediction study. Of course, I will
discuss several hypotheses and outline the related research questions.
2.1 What are Lexical Ambiguity and Polysemy?
In this sense prediction study, the main aim is to explore all possible senses for
undefined words; usually, these words have two or more different senses, have more
ambiguous interpretations, and have more polysemous explanations. In other
words, these words are regarded as lexically ambiguous or polysemous. However,
concerning semantic knowledge, there are also some differences between lexical
ambiguity and polysemy. Therefore, it is necessary to define both lexical ambiguity
and polysemy in order to determine how undefined words are classified. Moreover,
in this chapter, I will determine whether the four target words in this study are
classified as lexically ambiguous or are classified as polysemous.
2.1.1 Lexical Ambiguity
In lexical semantics, computational linguistics, and psycholinguistics, the issue of
lexical ambiguity is often discussed. Many scholars talk about polysemy and lexical
ambiguity in their studies because they are related concepts. However, they are
different when it comes to vague words versus ambiguous words and polysemous
words versus homonymous words.
© Springer-Verlag Berlin Heidelberg 2015
J.-F. Hong, Verb Sense Discovery in Mandarin Chinese—A Corpus based
Knowledge-Intensive Approach, DOI 10.1007/978-3-662-44556-3_2
2 Previous Researches on Lexical Ambiguity and Polysemy
Lexical ambiguity and polysemy both indicate vague, unclear, and indefinite
senses; that is to say, lexically ambiguous words and polysemous words can refer to
more than two senses at the same time. Because they are so similar, it is necessary to
define lexical ambiguity and polysemy as accurately as possible. Therefore, I will
discuss lexical ambiguity in this section and will discuss polysemy in the next section.
According to interpretation and comprehension, lexical ambiguity is the property
of being ambiguous; that is, a word, term, notation, sign, symbol, phrase, sentence,
or any other form used for communication is called ambiguous if it can be interpreted in more than one way. Lexical ambiguity (bank) is different from vagueness
(aunt), which arises when the boundaries of meaning are indistinct. Lexical
ambiguity is context-dependent: the same linguistic item (be it a word, phrase, or
sentence) may be ambiguous in one context and unambiguous in another context.
For a word, lexical ambiguity typically refers to an unclear choice between different
definitions as may be found in a dictionary. A sentence, however, may be
ambiguous due to different ways of parsing the same sequence of words.
Lexical ambiguity is a linguistic term for a word’s capacity to carry two or more
obviously different meanings, for example, bank. The word “bank” has several
distinct lexical definitions, including “financial institution” and “edge of a river.”
The context in which a lexically ambiguous word is used often makes evident
which of the meanings is intended. Therefore, if someone uses a multi-defined
word, it is sometimes necessary to clarify the context by elaborating on the specific
intended meaning (in which case, a less ambiguous term should have been used).
Lexical ambiguity arises when a word or concept has an inherently diffuse meaning
based on widespread or informal usage. This is often the case, for example, with
idiomatic expressions, whose definitions are rarely if ever well defined and are
presented in the context of a larger argument that invites a conclusion.
Lexical ambiguity is one of the most difficult problems in language processing
studies and thus, not surprisingly, it is at the core of lexical semantics research.
Concerning the distinction of lexical ambiguity, Weinreich’s (1964) distinction
between contrastive lexical ambiguity and complementary ambiguity was illustrative to this point. Contrastive lexical ambiguity is the situation where a lexical item
is associated with at least two distinct and unrelated meanings while complementary
ambiguity must be distinguished in a full semantic description: a purely formal
analysis, without reference to the substance.
In some modern linguistic and literary theories, it is argued that all signs are
polysemous, and the term has been extended to larger units, including entire literary
works. In WordNet, the definition of a lexically ambiguous word is the ambiguity
of an individual word or phrase that can be used (in different contexts) to express
two or more different meanings. This definition is also used for polysemy. In other
words, WordNet research team members regard polysemy and lexical ambiguity as
When determining the sense of a word, it is useful to distinguish the three stages
of processing lexical ambiguity: (1) decoding the input and matching it with a
lexically ambiguous word; (2) accessing the information about the ambiguous word;
and (3) integrating the information with the preceding context (Cottrell 1984).
2.1 What are Lexical Ambiguity and Polysemy?
Therefore, when defining lexically ambiguous senses, it is important to keep in mind
that (1) senses are represented as sets of necessary and sufficient conditions that fully
capture the conceptual content conveyed by words; (2) there are as many particular
senses for a word as there are differences in these conditions; and (3) senses can be
represented independently of the context in which they occur.
2.1.2 Polysemy
Polysemy is also a linguistic term for words with two or more meanings, usually
multiple and related meanings for a word or words. The words polysemy and
polysemous are defined as having or characterized by many meanings; the existence
of several meanings for a single word or phrase. When polysemous words are
discussed, homonymous words are likely to be discussed at the same time. However, polysemous words present different related meanings while homonymous
words present unrelated meanings.
Since the vague concept of relatedness is one test for polysemy, judgments of
polysemy can be very difficult to make. Because applying pre-existing words to
new situations is a natural process of language change, looking at a word’s etymology is helpful in determining polysemy but this is not the only solution; as
meanings become lost in etymology, what once was a useful distinction of meaning
may no longer be so. Some apparently unrelated words share a common historical
origin; however, etymology is not an infallible test for polysemy, and dictionary
writers often defer to speakers’ intuitions to judge polysemy in cases where it
contradicts etymology. Many words in Chinese are polysemous. For example the
verb 打 (da3 “hit”) can mean 打手臂 (da3 shou3bi4 “hit the back of a hand”), 輪胎
打氣 (lun2tai1 da3qi4 “pump gas into tire”), 打針 (da3 zhen1 “inject”), 把碗打破
(ba3 wan3 da3 po4 “break a bowl”), etc. (Hong et al. 2007, 2008).
There are several tests for polysemy, but one in particular is zeugma: if one word
seems to exhibit define when applied in different contexts, it is likely that the
contexts bring out different polysemy of the same word. If two senses of the same
word do not seem to match [e.g., 打 (da3 “hit”)], yet they seem related, then it is
likely that they are polysemous. The fact that this test again depends on speakers’
judgments about relatedness, however, means that this test for polysemy is not
infallible but rather is merely a helpful conceptual aid.
The study of polysemy, the multiplicity of meanings of words, has a long history
in the philosophy of language, linguistics, psychology, and literature (Ravin and
Leacock 2000). Ravin and Leacock (2000) pointed out three major approaches to
semantics represented in polysemy study: (1) the Classical Approach (e.g., Goddard
2000); (2) the Prototypical Approach (e.g., Fillmore and Atkins 2000); and (3) the
Relational Approach (e.g., Fellbaum 2000).
While the classical theories emphasizes definitions and related meaning to truth
conditions, possible words, and states of affairs, prototypical approaches emphasizes
meaning as part of a larger cognitive system and relates it to mental representations,
2 Previous Researches on Lexical Ambiguity and Polysemy
cognitive models, and bodily experiences. It is problematic to represent polysemy
within a relational framework, as polysemous word senses can be very distant from
each other in the semantic network’s conceptual space.
In addition, Geeraerts (1993) emphasized the importance of context when
determining the predictions of each of his tests, as he demonstrated that context
alters the senses of the words found in it. This emphasis on context is common to all
lexical ambiguity studies. In the above section on lexical ambiguity, it was mentioned that senses could be represented independently of the context in which they
occur. However, it is very important to focus on context for both lexical ambiguity
studies and polysemy studies.
In general, when talking about lexically ambiguous words or polysemous words,
word sense disambiguation (WSD) also should be taken into consideration. Karov
and Edelman (1998) pointed out the typical construct of WSD as follows:
Word sense disambiguation (WSD) is the problem of assigning a sense to an ambiguous
word, using its context. We assume that different senses of a word correspond to different
entries in its dictionary definition. For example, suit has two senses listed in a dictionary:
‘an action in court,’ and ‘suit of clothes.’ Given the sentence The union’s lawyers are
reviewing the suit, we would like the system to decide automatically that suit is used there
in its court-related sense (we assume that the part of speech of the polysemous word is
In other words, if researchers would like to decide the correct sense for polysemous words automatically based on context, they will refer to computer applications. According to Yael and Leacock (2000), computer applications that handled
the content of natural language texts need to come to terms with polysemy. They
consider that the study of polysemy in computational linguistics addresses the
problem of how to map expressions to their intended meanings automatically. As a
matter of course, it is very important to employ computer applications in polysemy
studies. More extendedly, corpus-based approaches have been used and machinereadable corpora have also existed.
Finally, using computer applications to deal with word sense disambiguation in
polysemy studies is reasonable because corpus-based approaches can provide statistical corpus analyses and the machine-readable corpora can provide large-scale
data. In addition, the three approaches (the Classical Approach, the Prototypical
Approach, and the Relational Approach) that Ravin and Leacock (2000) pointed out
are appropriate when dealing with different problems in polysemy studies.
2.1.3 The Relationship Between Lexical Ambiguity
and Polysemy
In this study, I choose four undefined target words with the intent of finding their
correct senses and assigning their appropriate senses based on different contexts. If
a word has more than two senses at the same time, then it is usually called a
lexically ambiguous word or a polysemous word. However, “lexical ambiguity” or
2.1 What are Lexical Ambiguity and Polysemy?
“polysemy” is presented in several different related researches as referring to the
same target. Even though WordNet research team members regard lexical ambiguity and polysemy as synonymous, lexical ambiguity and polysemy also can be
used in different contexts to represent two or more different meanings. It is very
difficult to differentiate lexically ambiguous words and polysemous words because
they have common points: more than two senses, vague senses, related senses, and
extended senses at the same time.
In fact, lexical ambiguity and polysemy are concepts used in several perspectives, such as Information Retrieval, computational approaches, natural language
processing (NLP), artificial intelligence, semantics, pragmatics, discourse, psycholinguistics, and neuropsychology. That is to say, lexical ambiguity and polysemy are very similar among different fields.
The main aim of this sense prediction study is to predict all possible senses for
the four target words—chi1 “eat”, wan2 “play”, huan4 “change”, and shao1 “burn”
rather than to disambiguate word senses. Because I will be using a large-scale
corpus as my empirical data in this study, firstly, I will extract the collocation words
of the four target words and cluster related collocation words in order to obtain
words that have the same senses and then cluster them in the same cluster. I will
then attempt to use the collocation words as intermediaries in order to predict all
possible senses and to examine all sentences of the four target words. In doing so, I
can obtain and predict all possible senses for the four target words from these
sentences. In the previous sections, I defined lexical ambiguity and polysemy and
introduced the following differences: (1) it is necessary to disambiguate word
senses in polysemy studies; and (2) it is necessary to divide word senses in lexical
ambiguity studies. Since the main goal of this study is to predict the correct senses
for the four target words chosen by dividing word senses rather than word sense
disambiguation in this study.
In sum, the main work of this study is to explore, predict, and obtain all possible
senses from all sentences for the four target words—chi1 “eat”, wan2 “play”, huan4
“change”, and shao1 “burn”—rather than disambiguate all possible word senses
based on the context of the target words. Finally, I will regard the four target words
as lexically ambiguous words and therefore will use the linguistic terms “lexical
ambiguity” and “lexically ambiguous words” throughout the remainder of this
sense prediction study.
2.2 Corpus-Based and Computational Model
2.2.1 Review of Previous Studies
Computational programming systems are designed to determine the appropriate
senses of words as they appear in linguistic contexts. Therefore, it is necessary to
review previous corpus-based and computational studies and discuss their models
and approaches.
2 Previous Researches on Lexical Ambiguity and Polysemy
The focus of this study aims to look for a unified analysis of lexical ambiguity,
as the problem of lexical ambiguity often poses theoretical and computational
problems in lexical semantic studies (cf. Ravin and Leacock 2002). Several previous studies concerning lexical ambiguity are well-cited in the literatures. However, the focus of these studies is nearly all on verbs rather than nouns. These
studies include mental processing comprehension (Ekaterini 2002), lexicon and
WordNet interpretations (McRoy 1992; Heiko 2002; Wu 2003; Buscaldi et al.
2007), context-based analysis (Jos’e et al. 2005; Cyma 2006; Wong et al. 2006),
information retrieval and machine translation (Li et al. 2000; Jos’e et al. 2005;
Wong et al. 2006; Zhou et al. 2006; Buscaldi et al. 2007), and lexical semantic
knowledge representation and frame-based approach (Bolette 1997; Lien 2000; Hsu
and Liu 2004; Liu et al. 2005).
In the case of the previously mentioned related lexical ambiguity studies in this
chapter (i.e., studies based on the corpus-based and computational perspective,
psycholinguistics perspective, and neurolinguistics perspective), I will divide these
lexical ambiguity studies into three different categories, list previous studies in these
three categories, and point out their significance, as shown in Table 2.1. I will focus
on discussing previous corpus-based and computational studies and point out the
gaps of these previous studies.
Veronis and Ide (1990) described a means for automatically building very large
neural networks (VLNNs) from definition texts in machine-readable dictionaries
and demonstrated the use of these networks for WSD. In their model, words were
complex units. Each word in the input was represented by a word node connected
by excitatory links to sense nodes, representing the different possible senses for that
word in the Collins English Dictionary. However, as they noted several improvements can further be made: (1) the parts of speech (POS) for input words and words
in definitions can be used to extract only the correct lemmas from the dictionary; (2)
the frequency of use for particular senses of each word can be used to help choose
among competing senses; and (3) additional knowledge can be extracted from other
dictionaries and thesauruses.
An up-to-date sampling of a wide range of methods can be found in a special
issue of Computational Linguistics on WSD (Philip and Yarowsky 2000). Annotated data has since facilitated recent advances in POS-tagging, parsing, and other
language processing sub-problems. They also presented a substantial exploration of
the relationship between monolingual sense inventories and translation distinctions
across languages.
Regarding Canas et al.’s (2003) study, they proposed using an algorithm to (a)
enhance the “understanding of the concept map by modules in the CmapTools
software that aide the user during map construction”; and (b) sort the meanings of a
word selected from a concept map according to their relevance within the map
when the user navigates through WordNet’s hierarchies, searching for more
appropriate terms. They presented the possibility of using an algorithm that exploits
WordNet to disambiguate the sense of a word that is part of a concept or linking
phrase in a concept map. The results shown were encouraging and suggest more
research should be done to improve the algorithm.
2.2 Corpus-Based and Computational Model
Table 2.1 Three categories of lexical ambiguity studies
Previous related studies
Significant points
Veronis and Ide (1990)
Philip and Yarowsky (2000)
Canas et al. (2003)
Ganesh and Prithviraj (2004)
Ker and Chen (2004)
Moldovan and Novischi
Chen et al. (2005)
Zhang et al. (2005)
Martinez et al. (2006)
Xue et al. (2006)
Peng et al. (2007)
Kipper et al. (2008)
Chen and Palmer (2009)
Pitler et al. (2009)
Tabossi and Zardon (1993)
Li and Yip (1996, 1998)
Li (1998)
Ahrens (1998, 2001, 2006)
Lin and Ahrens (2000)
1. Uses the corpus-based approach
2. An adaptive system
3. Based on context
4. Divides the sense of lexically ambiguous words
5. Finds the possible senses of a word
Gunter et al. (2003)
Li et al. (2004)
Elston-Guttler et al. (2006)
Mason and Just (2007)
Zempleni et al. (2007)
1. Experimentally-based
2. Determines literal bias meanings or
metaphorical bias meanings
3. Context influences lexical access
4. Conceptual domains and the linguistic
1. Takes lexically ambiguous words to
examine comprehensions of different
2. Processes ambiguous words that can
occur both as nouns and as verbs
3. Examines lexical ambiguity comprehension in order to determine the meanings of literal bias or metaphoric bias
Ganesh and Prithviraj (2004) introduced the notion of soft WSD, which states
that given a word, the sense disambiguation system should not commit to a particular sense but, rather, should commit to a set of senses that are not necessarily
orthogonal or mutually exclusive. In their work, WordNet gave multiple senses for
a word, which were related and which helped connect other words in the text. They
defined soft WSD as the process of enumerating the senses of a word in a ranked
order. This could be an end in itself or an interim process in an IR task, such as
question answering. They also found a Bayesian belief network (BBN), a natural
structure to encode such combined knowledge from WordNet corpus for training.
Ker and Chen (2004) described a general framework for adaptive WSD. Three
issues must be addressed in a lexicalized statistical WSD model: (1) data sparseness; (2) lack of abstraction; and (3) static learning. They also mentioned that an
adaptive system is superior in two ways to static word-based models trained on a
2 Previous Researches on Lexical Ambiguity and Polysemy
corpus. Through this learning strategy, an initial knowledge set for WSD was first
built based on the sense definition in training data.
Moldovan and Novischi (2004) showed how lexical chains and other applications could be built on this richly connected WordNet. They used the senses of the
words and defined them in WordNet. In order to overcome the data sparsity
problem, they relied on a set of methods that showed that disambiguation classes of
words share a common property. A suite of heuristical methods was presented for
the disambiguation of WordNet glosses. Moldovan and Novischi have used lexical
chains successfully to link question keywords with answer texts, providing axioms
to a Question-Answering logic prover.
In a different approach, the contexts that include ambiguous words are converted
into vectors by means of a second-order context method, and these context vectors
are then clustered by the k-means clustering algorithm (Chen et al. 2005). The
k-means clustering approach is an important method for data mining and knowledge discovery, as it has the characteristics of simplicity and fast convergence.
Zhang et al. (2005) proposed a corpus-based Chinese WSD approach using
HowNet. The approach is based on the following observation: The different senses
of a polysemous word tend to appear in cognizably different contexts. They
described a method that performs Chinese WSD by combining lexical co-occurrence knowledge, semantic knowledge, and domain knowledge. By this approach,
the experimental results showed that the method is very promising for Chinese
WSD in that study.
Martinez et al. (2006) observed that each algorithm, based on Leacock et al.
(1998), performed better for different types of words and each of them failed for
particular words. They observed a similar performance in preliminary experiments
when using an ML method or applying a heuristic on the different factors. They also
built a disambiguation algorithm that can be explained in four steps. The results
showed that the new method clearly outperforms the monosemous relatives in that
dataset. However, they also noticed that this improvement does not happen for all
the words in the set.
Concerning computational systems, Xue et al. (2006) devised a WSD system to
analyze ten highly polysemous verbs in Chinese. They compared the features they
used for Chinese with those used in a similar English WSD system. In that study,
they demonstrated that rich linguistic features, specifically features based on syntactic and semantic role information, are useful for the WSD of Chinese verbs.
Peng et al. (2007) mentioned that collocation was a combination of words that
has a certain tendency to be used together—and this was used widely to attack the
WSD task—and word classes were often used to alleviate the data sparseness in
NLP. They claimed that the algorithm of extending the collocation list that was
constructed from the sense-tagged corpus was quite straightforward. In their
experiments, the precision was proportional to the number of word classes. The
results of these experiments have shown that the average F-measure improved to
70.81 % compared to 54.02 % of the baseline system where the word classes were
not considered, although the precision decreased slightly.
2.2 Corpus-Based and Computational Model
Several scholars are still devoted to related works of the sense prediction study
or sense distinction performance study, such as Kipper et al. (2008), Pitler et al.
(2009), and Chen and Palmer (2009). Kipper et al. (2008) mentioned that lexical
classifications have proved useful in supporting various NLP tasks and some
information of VerbNet (VN). VerbNet is an extensive on-line lexicon for English
verbs, providing detailed syntactic-semantic descriptions and a hierarchical domainindependent, broad-coverage verb lexicon with mappings to several widely used
verb resources. They integrated two extensions into VN and incorporated the new
classes into VN. Therefore, these steps were syntactic descriptions, thematic roles,
and semantic descriptions of classes, such as entirely novel classes, novel subclasses, and classes where restructuring was necessary. Many uses of verb classes in
VN were being attested in a variety of applications, such as automatic verb
acquisition, semantic role labeling, robust semantic parsing, word sense disambiguation, building conceptual graphs, and creating a unified lexical resource for
knowledge extraction.
In another recent automatic sense prediction study, Pitler et al. (2009) worked
with a corpus of implicit relations present in newspaper text and reported results on
a test set. They used several linguistically informed features: polarity tags, Levin
verb classes, length of verb phrases, modality, context and lexical features, and used
the Penn Discourse Treebank (PDTB). They examined the most informative word
pair features and found that they were not the semantically related pairs that
researchers had hoped. In order to identify features useful for classifying comparison and other relations, they chose a random sample of 5,000 examples for contrast
and 5,000 other relations. They used experiments to demonstrate that features
developed to capture word polarity, verb classes, and orientation and found that
several lexical features were strong indicators of this type of discourse relation.
In the case of Chen and Palmer (2009), they discussed a high-performance,
broad-coverage supervised WSD system for English verbs that used linguistically
motivated features and a smoothed maximum entropy machine-learning model. In
their work, there were three major aspects: (1) developing a high-performance
WSD system for English verbs by using linguistically motivated features; (2)
applying this system to the first large-scale annotation effort aimed specifically at
providing suitable training data for high-performance WSD, followed by discussion
and analysis of these results; and (3) discussing potential future research areas for
large-scale, high-performance supervised WSD. In fact, their analysis showed that
using linguistically motivated features, such as semantic features, helped to relieve
the data sparseness problem. In addition, their experimental results on the larger set
suggested several areas they can explore in the future for improving high-performance WSD.
Some related previous studies were involved WSD in the sense prediction, I
consider that word sense induction (WSI) maybe more related for this sense prediction study. Navigli (2009) mentioned that a different approach to the induction of
word senses consisted of word clustering techniques, that was, methods which
aimed at clustering words which were semantically similar and could thus convey a
2 Previous Researches on Lexical Ambiguity and Polysemy
specific meaning. Navigli (2009) also mentioned that word sense induction was
performed with high precision (recall varies depending on part of speech and
In addition to the above previous studies, I have also investigated some representative studies concerning lexical ambiguity in lexical semantics. These include
Lexical Semantics and Meaning in Language (Cruse 1986, 2004), WordNet
(Fellbaum 1998), and Generative Lexicon (Pustejovsky 1991, 1995). From these
previous studies, I observed that lexically ambiguous word senses might include
several cases illustrating the relation of indefiniteness, in which the significant part
is more predominant than the overlapping semantic element.
In Pustejovsky’s (1995) generative lexicon study, especially, he discussed the
logical problem of polysemy and pointed out two types of ambiguity—contrastive
ambiguity and complementary polysemy—by following Weinreich (1964). Concerning contrastive ambiguity, Pustejovsky mentioned that given the current representational techniques and strategies for differentiating word senses, there would
appear to be no reason to make a logical distinction between these two types of
ambiguity. A dictionary called a Sense Enumeration Lexicon (SEL) was introduced,
and it appeared at first to handle adequately the sense differentiation for both
ambiguity types. From the theoretical perspective, the major problems posed by
contrastive ambiguity involved issues of discourse inference and the correct integration of contextual information into processing. Therefore, Pustejovsky brought
up the elementary lexical semantic theory and considered that the major part of
semantic research had been on logical form and the mapping from a sentence-level
syntactic representation to a logical representation language. In addition, regarding
the Sense Enumeration Lexicon, he characterized it directly as follows:
A lexicon L is a Sense Enumeration Lexicon if and only if for every word w in L, having
multiple senses s1, …, sn, associated with that word, then the lexical entries expressing
these senses are stored as {ws1, …, wsn}.
As in the example bank in above I mentioned, two contrastive senses could be
listed in a straightforward fashion as shown in (2.1) and (2.2), using a fairly
standard lexical data structure of category type (CAT) and a basic specification of
the genus term (GENUS), which locates the concept within the taxonomic structure
of the dictionary
@ CAT ¼ count noun
GENIUS ¼ financial institution
@ CAT ¼ count noun A
GENIUS ¼ store
All possible selectional requirements of verbs are defined from the features or
types as the genus terms, and disambiguation would appear to be merely the process
2.2 Corpus-Based and Computational Model
of correctly matching the features of functor and arguments from the available set of
lexical entries.
Although this approach was taken by many researchers within both theoretical
and computational traditions, Pustejovsky presented three arguments against using
the SEL as a model of lexical semantics: (1) “The Creative Use of Words”—the
SEL cannot capture the full range of word usages; (2) “Permeability of Word
Senses”—the SEL cannot capture the relationship between senses; and (3) “Difference in Syntactic Forms”—the SEL cannot allow senses to have an adequate
range of syntactic realizations.
These arguments present problems in defining a set of features or types for
contrastive senses of the verbs in the Sense Enumeration Lexicon. It is necessary to
improve this approach and then make a useful model to deal with lexically
ambiguous words, which is the aim of this sense prediction study. I will also use the
corpus-based and computational approach but with two different strategies—character similarity clustering analysis and concept similarity clustering analysis. I
expect the results to be better than the results using the Sense Enumeration Lexicon.
2.2.2 Gap of Previous Studies
Overall, regarding these previous corpus-based and computational studies, these
scholars proposed corpus-based, algorithm, automatically computational programming system, and collocation approaches to analyze sense prediction studies or
WSD studies. Moreover, they also recommended that using large-scale corpus,
context, semantic features, and concepts could achieve high performance for sense
prediction studies. In the above studies, I found that they generally employed only
one corpus in their studies, which resulted in less information of lexical ambiguity
for their sense prediction studies; they also did not combine the various approaches
Focusing on some previous studies of them, I consider that specific research
gaps existed and these research gaps were easily observed. For example, they were
observed in Ker and Chen (2004), Chen et al. (2005) and Peng et al. (2007).
Ker and Chen (2004) mentioned that the first step of their study was to construct
an initial knowledge from training corpus. However, they did not point out how
many training data which they selected can let them obtain better performance
under the adaptive sense disambiguation approach. In this study, I will point out the
number of predicting clusters as my default target for the four target words to
present the best results.
In Chen et al. (2005), they didn’t explain what was about second-order context
and why this approach could provide more information about word senses in
contexts. In addition, they mentioned that the whole process was completed automatically, so a sense-labeled corpus was not need. In my study, I will not only
predict all possible senses of the four target words by automatically computational
programming, but I will also examine these clusters whether they can be predicted
2 Previous Researches on Lexical Ambiguity and Polysemy
appropriate senses in manual. In addition, I will use senses divisions in Chinese
Wordnet and Xiandai Hanyu Cidian (Xian Han) to estimate the evaluations of the
four target words by my own intuition.
Peng et al. (2007) took the target verb 吃 chi1 “eat” as the illustration and
selected the number of word classes. In their study, they only talked about the
concrete objects for 吃 chi1 “eat” but no abstract objects for 吃 chi1 “eat”. In my
corpus-based and computational analysis, I will predict physical senses and metaphorical senses of the four target words.
In consideration of several research gaps presented by these previous studies,
this study utilized four corpora in order to obtain richer information, automatically
computational programming to gather related collocation words of the four target
words—chi1 “eat”, wan2 “play”, huan4 “change”, and shao1 “burn”, and used
HowNet in an attempt to identify their semantic features and elements. In addition,
this study adopted the same morpheme contrast and concept contrasts by automatically computational programming in the corpus-based and computational
approach in order to divide the sense clusters of the four target words.
2.3 Hypotheses and Research Questions
With respect to the sense prediction study of lexical ambiguity, there are three
hypotheses in this study. Lexical ambiguity means some words have multiple
meanings or senses (Moldovan and Novischi 2004). In the SEL model, although
lexically ambiguous words list all possible selectional requirements that are associated with those words and then lexically ambiguous words express these senses,
the SEL model can not capture the full range of word usages. Therefore, the first
hypothesis is that words with similar morpheme-character components and concept
elements are similar in sense. I will follow Fujii and Croft (1993) to observe
character similarity and refer to Li et al. (2003) and Dai et al. (2008) to explore
concept similarity via HowNet.
Peng et al. (2007) mentioned that a corpus was divided into five equal parts
which one part was used as the test corpus and the collocation list was constructed
from the other four parts of corpus. In this study, the second hypothesis is that
different corpora with particular functions which provide different lexical knowledge bases. I will use Chinese Gigaword Corpus to select related collocations for
the four target words; I will use HowNet to assign all possible concepts to
ambiguous senses of the four target words; I will use Chinese Wordnet to estimate
the evaluations for the four target words and I will also use Xian Han to estimate the
evaluations for the four target words.
According to Ahrens’ studies (1998, 2001, 2006), she considered that sentential
context and meaning frequency could influence the lexical ambiguity resolution and
access. Owing to all possible clusters of all collocation words of the four target
words being selected from the character similarity clustering analysis by examining
their contexts, I consider these collocation words, which are stimuli for the
2.3 Hypotheses and Research Questions
experimental evaluations, to have been identified by their frequencies and their
senses in different contexts. Hence, the third hypothesis is that in the off-line
multiple-choice task, subject uses conceptual difference to identify the choice. In
this study, I will use an off-line multiple-choice task involving experimental evaluation in order to examine which concept of one selected word/item is obviously
different from the concepts of the other three words/items. Stimuli are selected from
the character similarity clustering analysis by examining their contexts and they are
controlled the frequencies. And then, I can demonstrate other approaches which can
verify the analysis of the corpus-based and computational approach.
For this reason, there are three research questions in this study: (1) How do I
predict the word senses of a lexically ambiguous word in order to present different
interpretations in different contexts or domains? (2) How do I use more than two
corpora as the database to support this word sense prediction study? (3) Can I use
other approaches to verify the analysis of the corpus-based and computational
approach for this word sense prediction study?
I will make use of the lexical semantics, lexical features, concepts, and collocation words to examine these research questions using Chinese Gigaword Corpus,
HowNet, Chinese Wordnet, and Xiandai Hanyu Cidian. Therefore, I will attempt to
utilize the corpus-driven linguistic approach as my main method for this sense
prediction study.

Report this document