AN EMPIRICAL STUDY OF WORD SENSE DISAMBIGUATION

International Journal on Natural Language Computing (IJNLC) Vol. 5, No.5, October 2016
DOI: 10.5121/ijnlc.2016.5503 29
AN EMPIRICAL STUDY OF WORD SENSE
DISAMBIGUATION
M.Srinivas1
and Dr. B.Padmaja Rani 2
1
Department of Computer Science & Engineering, Geethanjali College of
Engineering & Technology, Hyderabad, India
2
Department of Computer Science & Engineering, JNTU-CEH, Hyderabad, India
ABSTRACT
Word Sense Disambiguation (WSD) is an important area which has an impact on improving the
performance of applications of computational linguistics such as machine translation, information
retrieval, text summarization, question answering systems, etc. We have presented a brief history of WSD,
discussed the Supervised, Unsupervised, and Knowledge-based approaches for WSD. Though many WSD
algorithms exist, we have considered optimal and portable WSD algorithms as most appropriate since they
can be embedded easily in applications of computational linguistics. This paper will also provide an idea of
some of the WSD algorithms and their performances, which compares and assess the need of the word
sense disambiguation.
KEYWORDS
Supervised, unsupervised, knowledge-based, WordNet, word sense disambiguation
1. INTRODUCTION
The creatures are using the language as their communication media. Through language
information can be exchanged among the races. Verbal communication involves alphabets,
words, sentences etc. In almost all natural languages, there are words having different meanings
depending on the context. Those words are known as polysemous words making verbal
communication ambiguous. Fortunately, human beings resolve the ambiguity instantly depending
on the context with lot of ease. But, machines find it as a very difficult problem. This involves
processing unstructured textual information to build the appropriate data structures. We determine
the most appropriate meaning through analysing the data structures thoroughly. This is known as
Word Sense Disambiguation, a common problem in Natural Language Processing (NLP).
This paper is organized as follows: In section II a brief description of WSD, presenting main
approaches knowledge based, supervised, and unsupervised disambiguation in sections III, IV, V
respectively. In section VI we elaborated on some evaluation measures for assessing WSD
systems and conclusion in section VII followed by the references.
2. WORD SENSE DISAMBIGUATION
Word sense Disambiguation is the process of identifying the correct sense of a word that has
several meanings in the context in a computational paradigm. Machine translation is one of the
most former on growing research topic in computational linguistics. The problem WSD is as
complex as most of the difficult problems in Artificial Intelligence and hence it is deemed as an
AI complete problem. In 1940’s WSD was developed as discrete field in computational

30
linguistics to help the research in machine translation. In 1950’s Weaver identified that context is
crucial and hence, statistical semantic studies have been undertaken as a necessary primary step.
The automatic disambiguation of word senses has been given an utmost priority from the earliest
days of computer treatment of languages in the 1950’s. WSD depends heavily on knowledge
sources like dictionaries, thesauri, ontology’s, collocations, WordNet etc. WSD can be described
as a method of providing the most appropriate sense to all or some words in the text where T is a
sequence of words (w1, w2, ...., wn). It is called as All-words WSD when it attempts to
disambiguate all words in a text such as nouns verbs, adjectives, adverbs¸ etc. Otherwise Targeted
WSD as it disambiguates some restricted words only. It consists of mainly discovering the
mapping M from words to senses such that M(k) ⊆SensesD(wk ) where M(k) is the subset of
senses of wk which are appropriate in the text T and SensesD(wk) is the set of senses in dictionary
D for word wk. The mapping M can assign multiple senses to wk belonging to T but finally the
most appropriate sense is selected. Hence, WSD can be seen as a classification task where word
senses form the classes and a method classifies each occurrence of the word to multiple classes by
exploiting information available from the context and external knowledge sources such as
dictionary, thesauri, ontology’s, collocations, wordnet, unlabelled or annotated sense corpora.
The input text is pre processed to build a structured format suitable for our WSD system. It
consists of the following steps in sequence:
a. Tokenization - Dividing the text into basic units (tokens) called as words.
b. Part -of-Speech Tagging - Determining the appropriate grammatical category for each word.
c. Lemmatization - Performs morphological analysis to provide the root words.
d. Chunking - Partitioning the text in syntactically correlated parts.
e. Parsing- Provides the parse tree of sentence structure.
Following the above pre processing, each word is represented as a feature vector making the
assignment of the appropriate sense easy by the WSD system.
3. KNOWLEDGE BASED APPROACHES
The objective of knowledge based approaches is to make extensive use of knowledge resources
like WordNet to determine the senses of words in context. These methods have lower
performance than their counterpart supervised methods, but they can be applicable to a wider
range. The knowledge based approaches to word sense disambiguation started initially in the year
1979 and 1980, conducting experiments on extremely limited domains. The unavailability of
large-scale computational resources restrained scaling up these works due to difficulty in
performing proper evaluation, comparison and exploration of these methods especially when
applied to end-to-end applications.
We present the following essential knowledge-based techniques: the overlap of sense definitions,
selectional restrictions, and structural approaches. Many of these exploit information from
WordNet or other resources.
3.1. Overlap of Sense Definitions
It is a straightforward knowledge based approach that consists of calculating the word overlap
between the sense definitions of two or more target words as a primary step. This approach is
termed as gloss overlap or the LESK algorithm [1]. Given a two-word context (w1, w2), the senses
of target words having the highest overlap in their definitions are chosen as the correct ones. For
two words w1 and w2, we determine score for each pair of word senses S1 ∈Senses (w1 ) and S2
∈Senses(w2 ):
Score Lesk ( S1 , S2 ) =| gloss(S1 ) ∩ gloss(S2 ) |,

31
where gloss(Si) is the set of words in the textual definition of sense Si of wi. The senses with
maximum value for the above formula are assigned to the respective words. As it requires
exponential number of steps, a variant of Lesk is currently used which identifies the sense of a
word w whose textual definition has the highest overlap with the words in the context of w. For a
target word w, corresponding score is computed for each sense S of w as follow:
Score LeskVar (S) =| context (w) ∩ gloss(S) |,
where context (w) is the set of all content words in a context window around the target word w.
Lesk’s approach is very sensitive to the exact wording of definitions, so the absence of a
particular word can drastically change the results.
A measure of extended gloss overlap as discussed in [2], which expands the glosses of the words
being compared to include glosses of concepts that are known to be related through explicit
relations in the dictionary (e.g.,hypernymy, meronymy, pertainymy, etc.). The range of
relationships used to extend the glosses is a parameter, and can be selected from any combination
of WordNet relations.
For each sense S of a target word w, we estimate its score as
scoreExtLesk(S) = ∑ rel | context(w) ∩ gloss(S ) |,
SI
:S → SI
or S≡SI
where context(w) is the bag of all content words in a context window around the target word w
and gloss(S ) is the bag of words in the textual definition of a sense S which is either S itself or
related to S through a relation rel. The overlap scoring mechanism is also parameterized and can
be modified to take into account gloss length or to include function words as
scoreExtLesk(S) = ∑ rel | context(w) ∩ gloss(S ) |,
SI
:S → SI
or S≡SI
3.2. Selectional Preferences
It is one which exploits selectional preferences to constrain the number of meanings of a target
word occurring in context. Selectional preferences or restrictions restrict the semantic type that a
word sense imposes on the words with which it combines in sentences. For instance, the verb
drink expects an animate entity as subject and a potable entity as its direct object. We can
differentiate selectional restrictions and preferences in that the former rule out senses that violate
the constraint, whereas the latter tend to select those senses which better satisfy the requirements.
The simple way to learn selectional preferences is to find the semantic appropriateness of the
association provided by a word-to-word relation through frequency count. Given a pair of words
w1 and w2 and a syntactic relation R (e.g., subject-verb, verb-object, etc.), this method counts the
number of instances (R, w1, w2 ) in a corpus of parsed text, obtaining a measure Count(R, w1, w2)
as in [3]. Conditional probability, another measure of the semantic appropriateness of a word-to-
word relation, of word w1 given the other word w2 and the relation R:
P ( w1 /w2 , R) = Count(w1,w2,R)/Count( w2 , R)

32
WordNet helps to derive a mapping from words to conceptual classes. Various techniques have
been devised for measure of Selectional association [4, 5], to tree cut models using the minimum
description length [6, 7], hidden markov models [8], etc. Almost all these approaches exploit
large corpora and model the selectional preferences of predicates by combining observed
frequencies with knowledge about the semantic classes of their arguments. Finally
disambiguation is performed with different means based on the strength of a selectional
preference towards a certain conceptual class.
3.3. Structural Approaches
Multiple structural approaches have been developed to analyze and exploit the structure of the
concept network available through computational lexicons like WordNet. The recognition and
measurement of patterns, both in a local and a global context, can be accumulated in the field of
structural pattern recognition [9], which intend to classify data based on the structural
interrelationships of features. We discuss two prime approaches here: similarity-based and graph-
based methods.
3.3.1 Similarity Measures
We disambiguate a target word wi in a text T = (w1, . . . , wn ) by selecting the sense S of wi with
maximum value for the following sum:
Given a sense S of our target word wi, the formula adds the contribution of the most appropriate
sense of each context word wj = wi. The sense with the highest sum is selected. Similar
disambiguation strategies can be applied [10]. We now present well-known measures of semantic
similarity in the literature.
A simple metric depending on determining the shortest distance in WordNet between pairs of
word senses is introduced in [11]. The hypothesis is that, for given a pair of words w and wI
occurring in the same context, the most proper sense is determined through selecting the senses
that maximize the distance between w and wI.
It is defined as follows:
scoreRada (S w , S w
I
) = d (Sw , Sw
I
),
where d (Sw , Sw
I
) is the shortest distance between Sw and Sw
I
is obtained by counting the number
of edge of the shortest path in the lexicon network. The shortest path determined on the WordNet
taxonomy includes only hypernymy edges.
An approach in [12] is based on ascertaining that concepts deep in taxonomy are more closely
related to each another than those in the upper part of the same taxonomy. An edge in the
WordNet noun taxonomy is seen as a pair of two directed edges representing inverse relations
(e.g., kind-of and has-kind). The measure is defined as follows:
scoreSussna(Sw ,Sw
I )=
wR (sw ,Sw
I )+WR
−1 (Sw
I ,Sw)
2 D
where R is a relation, R−1
its inverse, D is the overall depth of the noun taxonomy, and each
relation edge is weighted based on the following formula:

33
where maxR and minR are a maximum and minimum weight that we want to assign to relation R
and nR (Sw) is the number of edges of type R emerging from Sw.
A similarity measure based on the distance between two senses Sw and Sw
I
is developed in [13].
They concentrated on hypernymy links and gradually incrementing the path length by the overall
depth D of the taxonomy:
The disadvantage of distance-based measures is not considering the density of concepts in a sub
tree whose root is a common ancestor. The variant of Conceptual density that computes the
density of senses of a word context is demonstrated in [14]. For a synset S, assuming m senses of
words to be disambiguated in its sub hierarchy a mean number of hyponyms per sense nhyp is
also taken into account. Then, the conceptual density S is determined as follows:
where descendants(S) is the total number of descendants in the noun hierarchy of S and 0.20 is a
smoothing exponent, whereas i ranges over all possible senses of words in the subhierarchy of S.
For all senses of nouns in the given context, the conceptual density S is determined for all of their
respective hypernyms. The set of sense choices is obtained from the synset associated with
highest conceptual density. The senses included in its sub hierarchy are considered as respective
words interpretation in context. The remaining senses of those words are removed from the
hierarchy, and the procedure is repeated for the other ambiguous words.
4. SUPERVISED DISAMBIGUATION
Machine-learning techniques are applied to construct a classifier from manually sense-annotated
data set in supervised word sense disambiguation approaches. The classifier or word expert
chooses a word at a time and accomplish a classification task in an attempt to determine the
appropriate sense for instance of that word. The training set consists of numerous examples
having the given target word manually tagged with a sense from the sense inventory of a
reference dictionary. This is used to learn the classifier. We discuss here well-known machine
learning techniques namely: Decision lists, Decision trees, Naive Bayes and Neural networks in
the following sections.
4.1. Decision Lists
A decision list [15] is an ordered set of rules for classifying test instances. It is a list of weighted
if-then –else rules, arranged in decreasing order of their score. A set of features are built using a
training set. Then rules of the form (F, S, T) where F denotes feature-value, S the sense and T its
score are obtained.
First, a feature vector representing the given target word w is obtained. Then, the decision list is
examined in order to find the feature with the maximum score matching the input vector. The
winning feature identifies the proper word sense for word w.

34
^S= argmaxSi
∈SensesD
(w)score(Si)
Yarowsky [16], modified the criteria to improve the performance. In this score of sense Si is
taken to be the maximum among the feature scores. The score of a feature f is determined through
the logarithm of the probability of sense Si given feature f divided by a value that is obtained
accumulating the probabilities of the other senses given feature f :
Smoothing and pruning are performed to remove zero counts and avoiding not reliable rules with
less weight respectively.
4.2. Decision Trees
In the decision tree classification rules are represented as a tree structure having the virtue of
recursively partitioning the training data set. Every internal node indicates a test on a feature
value. Branch followed specifies the outcome of the test. For a given word to assign the proper
sense, we start from the root performing a sequence of test following the corresponding branches
until a leaf node is reached, where sense assignment is done if it is not empty denoted by - .
Algorithms known for better performance for learning decision trees are C4.5 algorithm [17], ID3
algorithm [18]. When comparing the performance of decision tree obtained with the C4.5
algorithm with other machine learning algorithms, it is surfaced that the former results in less
performance. Though it has the advantage of representing the predictive model in a compact and
Human readable way, it suffers from issues like data sparseness because of features with more
number of values, unreliable predictions because of small training set, etc. An example of a
decision tree for WSD is described in Figure 1. For example, if the noun bank must be classified
in the sentence “I deposited money in my account” the tree is traversed starting from the root,
after following the branches, the choice of sense bank/ Finance is made.
Fig1.An example of a decision tree
4.3. Naive Bayes
A Naive Bayes classifier is a straightforward probabilistic classifier obtained through the use of
Bayes’ theorem. We determine the conditional probability of each sense Si of a word w for the set
of features fj in the context. Then we calculate the following formula for each sense.

35
The sense S maximizing the above is identified as the most appropriate sense in the context.
Smoothing is performed to eliminate the impact of zero counts as in [19]. But, this solution
causes probabilities to have value more than 1. We can address this problem through Back off or
interpolation strategies.
4.4. Neural Networks
A neural network [20] is an interconnected group of artificial neurons. It processes data using a
computational model based on a connectionist approach. The pairs consisting of (inut feature,
desired response) are accepted as input to the learning program.
The key objective is to using the input features to divide the training contexts into disjoint sets
associated with desired response. When the network is added new pairs, link weights are adjusted
properly so that outcome unit specifying the desired response has more activation than any other
outcome unit. Neural networks must be trained adequately so that the outcome of the unit
specifying the desired response is more than the outcome of any other unit for every training
example. Weights in the network are either positive or negative, accumulating the evidence
supporting or denying a sense choice Cottrell [21] used neural networks where nodes represent
words: the words activate the concepts to which they are semantically related and vice versa. The
activation of a node results in the activation of nodes to which it is connected through excitory
links and the deactivation of those to which it is connected through inhibitory links.
It is shown in [22, 23] that neural networks outperforms other supervised methods based on
experiments performed on few number of words. The complex interpretation of results, the
requirement of vast training data, and adjusting parameters like threshold, decay, etc are the
major drawbacks in using the neural networks.
5. UNSUPERVISED DISAMBIGUATION
The performance of supervised disambiguation approaches suffers from knowledge acquisition
bottleneck. But, unsupervised methods overcome the lack of large-scale resources annotated
with word senses manually by virtue as discussed in [24]. These are developed based on the
observation that word with a particular sense will always be surrounded by the similar words.
These approaches identify word senses from input text by dividing word occurrences. New
occurrences are clustered based on the induced clusters. Hence, these methods don not rely on
labelled training text and even not using the knowledge of machine-readable resources like
dictionaries, thesauri, WordNet, etc.
Unsupervised WSD performs word sense discrimination rather than assigning an appropriate
sense to a target word like in knowledge based or supervised approaches. It accomplishes it by
partitioning the occurrences of a word into multiple classes by identifying for any two
occurrences whether they correspond to the same sense or not as discussed in [25]. The
clusters obtained in these approaches may not be equivalent to the regular senses in a dictionary
sense inventory. This makes the evaluation more complex which may require human intervention
or indirect evaluation using the clusters in end-to-end applications.
However, sense discrimination and sense labelling are two sub problems of the WSD task as
discussed in [26], but are related closely. Hereafter, we discuss the key approaches to
unsupervised WSD, namely: methods based on context clustering, word clustering, and co
occurrence graphs.

36
5.1. Context Clustering
We have an unsupervised approach based on the notion of context clustering. Context vector is
used to represent each occurrence of a target word in a corpus. Then, these vectors are partitioned
into groups, each helps determining a sense of the target word.
A historical approach in this category is based on the idea of word space also known as vector
space, whose dimensions are words. A word w in a corpus is represented as a vector whose ith
component gives the total number of times that words wi and w appear within a fixed context.
Then, the similarity between two words v and w is computed geometrically by the cosine between
the corresponding vectors v and w:
where m is the number of features in each vector. A vector is determined for every word in a
corpus.
We obtain a cooccurrence matrix by combining the set of vectors for each word in the corpus. To
reduce the dimensionality of the resulting multidimensional space through singular value
decomposition (SVD), latent semantic analysis (LSA) is applied [27]. SVD identifies the major
axes of variation in the word space. The input to the dimensionality reduction is set of word
vectors in the high dimensional space and output is representation in a lower- dimensional space.
The dimensions corresponding to terms having similar meanings may be merged.
Following the dimensionality reduction, contextual similarity between two words is determined in
terms of the cosine between respective vectors. Next, our objective is clustering context vectors.
A context vector is constructed as the centroid of the vectors of the words appearing in the target
context that accounts for its semantic context.
Finally, sense discrimination is achieved by grouping the context vectors of a target word
applying a standard clustering algorithm. Context-group discrimination algorithm was proposed
by Schutze that categorize the occurrences of an ambiguous word into clusters of senses
depending on the contextual similarity between occurrences. Contextual similarity is calculated as
above. In this, clustering is performed with the Expectation Maximization algorithm based on the
variation of the probabilistic model [28]. A different clustering approach consists of
agglomerative clustering [29]. Initially, each instance consists of a cluster with single element.
Next, agglomerative clustering combines the most similar pair of clusters and repeated for
resulting less similar pairs until the corresponding threshold is not acceptable. Its performance
was assessed in the biomedical domain.
The difficulty in building the context vectors is that a vast amount of training data is necessary for
identifying a remarkable distribution of word co-occurrences. This problem can be resolved by
augmenting the feature vector of every word with the content words appearing in the glosses of
its senses [30]. A more subtle issue to be resolved is the fact that different context vectors are not
assigned to distinct word senses. In [31], it is resolved through a well trained supervised
classifier. Word senses are determined through the use of multilingual context vectors in [32].
Here, a word occurrence in a multilingual corpus is represented as a context vector that takes into
account every possible lexical translation of the target word w, whose value is 0 if the specific
occurrence of w cannot be translated properly, otherwise it is 1.

37
5.2. Word Clustering
In the previous method, first- or second-order context vectors are used to represent word senses.
A different approach for determining appropriate word senses is based on word clustering
techniques, whose objective is clustering words that are semantically similar and thereby
supporting a specific meaning.
A popular approach to word clustering involves the identification of words W=((w1, . . . , wk) that
are similar to a target word w0. For each wi its similarity with w0is measured using the
information content of their single features given by the syntactic dependencies like subject-verb,
verb-object, adjective-noun etc. The information content is high, when two words share
dependencies. Now, word clustering algorithm is applied to discriminate between the senses.
Suppose that W is the set of similar words ordered based on the degree of similarity to w0. A
similarity tree T is created which initially has a single node w0. Next, for each i∈{1, . . . , k}, wi
∈W a child is added to wj in the tree T where wj is the most similar word to wi among {w0 , . . . ,
w i−1 }. It is followed by pruning – deleting unnecessary nodes. Now each sub tree with root w0 is
considered as a distinct sense w0.
A different word clustering technique, known as Clustering By Committee (CBC) algorithm, was
discussed in [33]. In this, for every target word, a set of similar words is determined as above. To
determine the similarity, each word is represented as feature vector, where every feature is the
expression of syntactic context having the word.
For a given set of target words, first a similarity matrix S is constructed where Sij contains the pair
wise similarity between words wi and wj. Next step invokes a recursive procedure to determine set
of clusters known as committees, for each word in the given set of words E. One of the standard
clustering technique, average-link clustering is employed. In each step, words not covered by any
committee are discovered. Recursive attempts are done to identify more committees from residue
words. Finally, for each target word w ∈E, which is represented as a feature vector, is iteratively
allocated to its most similar cluster, depending on its similarity to the centroid of each committee.
After a word w is assigned to a committee c, the intersecting features between w and elements in
c are deleted from the representation of w, allowing the identification of less frequent senses of
the same word in subsequent iterations.
5.3. Cooccurrence Graphs
The word sense discrimination relying on graph-based approaches provides a different view,
which has been recently explored providing fairly good performance. Cooccurrence graph g-
(V,E), where V the set of vertices represents words in a text and edges E connect two words
having a syntactic relation, in the same paragraph, or in a larger context, is used in developing
graph-based approaches.
Cooccurrence graph built based on grammatical relations between words in context is described
in [34]. A local graph Gw is constructed around w for each target ambiguous word w. The graph
can be seen as a Markov Chain, after normalizing the adjacency matrix representing Gw. Then,
the Markov clustering algorithm [35] is applied to identify word senses, depending on an
expansion and an inflation step, focusing, respectively, at inspecting new more distant neighbours
and including other popular nodes.
Subsequently, an ad hoc approach known as HyperLex is proposed in [36]. In this, initially a
cooccurrence graph is constructed where vertices V is the set of words occurring in the paragraph
of a text corpus having the target word, and an edge between a pair of words is included in the

38
graph if they co-occur in the same paragraph as the first step. It is a weighted graph, where each
edge is given a weight based on the relative cooccurrence frequency of the two words connected
by the edge. So, for each edge (i, j) its weight wi j is calculated as
wi j = 1 − max{P (wi | wj ), P ( wj | wi)},
where P (wi | wj )= freq i j/ freq j , and freqi j is the frequency of cooccurrence of words wi and wj
and freqj is the frequency of wj within the text. Following the above, words which rarely occur
together are assigned weight close to 1, whereas words with high frequency of cooccurrence are
given weight close to zero.
The next step involves running an iterative algorithm on the cooccurrence graph built in the
above. In each iteration, we select as a hub, a node with highest relative degree in the graph. It
results in all its neighbours cannot be hubs. The algorithm ends when the relative frequency of the
word corresponding to the chosen hub is less than a fixed threshold. The entire set of hubs
selected so far, represent the senses of the word of interest. Then, these hubs are linked to the
target word and edges are assigned zero-weight Hubs are then linked to the target word with zero-
weight edges and the minimum spanning tree (MST) is built for the entire graph that is used to
disambiguate particular instances of our target word. Suppose that W = (w1 , w2 , . . . , wi , . . . ,
wn ) is a context having wi, our target word.
A score vector s is assigned to each wj ∈W (j = i), such that its kth
component sk represents the
contribution of the kth
hub as follows:
where d (hk , wj ) is the distance between root hub hk and node wj.
Next, all score vectors assigned to all wj ∈W (j ≠ i) are added and the hub which is associated
with the highest score is chosen as the most appropriate sense for w i.
An alternative graph-based algorithm for determining word senses PageRank is discussed in [37].
PageRank is popular algorithm devised for determining the ranking of web pages. Not only the
Google search engine depends more on page rank for its better performance, but also used in
many research areas where the prime objective is determining the importance of entities where
graph represents the relations among them. In its weighted formulation, the PageRank degree of a
vertex vi ∈V is determined using the following formula:
where v j→ v i represents the presence of an edge from vj to vi, d is a damping factor and wji is its
weight which models the probability of following a link to vi or randomly jumping to vi. The
above formula is recursive: the PageRank of each vertex is iteratively computed until
convergence.

39
6. EVALUATION METHODOLOGY
We discuss here the evaluation measures and baselines, borrowed from the field of information
retrieval, namely coverage, precision, and recall. However, the main objective of WSD is to
demonstrate that it improves the performance of applications such as machine translation,
information retrieval, question answering systems, etc. The evaluation of WSD that is part of
applications is called in vivo or end-to-end evaluation. We present this second kind of evaluation
in the following section.
6.1. Evaluation Measures
Let S = (w 1 , . . . , w n ) be a test set and T an “answer” function that associates with each word w
i ∈S the appropriate set of senses from the dictionary D (i.e., T(i) ⊆Senses D (wi )). Then, given
the sense assignments TI
(i) ∈Senses D (w i ) ∪{є} provided by an automatic WSD system (i ∈
{1, . . . , n}), then, coverage C is defined as the percentage of items in the test set for which the
system identify a sense assignment. So
C = # answers provided / # total answers to provide = | {i ∈{1, . . . , n} : TI
(i) =є } |/ n
Where є denotes that system fails in providing an answer for a specific word w(i.e., in
that case we assume that TI
(i) =є). The total number of answers is given by n =| S|.
The precision P of a system is calculated as the percentage of correct answers given by the
automatic system, that is:
P = # correct answers provided / # answers provided
= |{i ∈{1, . . . , n} : TI
(i) ∈A(i)} | / i ∈{1, . . . , n}: TI
(i)≠ є } |
Recall R is defined as the number of correct answers provided by the automatic system over the
total number of answers to be provided:
According to the above definitions, we have R ≤ P by default. When coverage is 100%, we have
P = R. In the WSD literature, recall is also called as accuracy, but these are two different
measures in the information retrieval and machine learning literature.
Finally, we have F1-measure or balanced F-score, computed as
F1 = 2PR/P + R
It is a measure calculating the weighted harmonic mean of precision and recall.
F1-score is useful to compare systems providing coverage less than 100%. If simple arithmetic
mean is used, we can easily build a system with P = 100% and almost-zero recall giving around
50% performance, but harmonic mean such as F1- score significantly reports lower performance
for low values of either precision or recall.

40
7. CONCLUSION
In this paper we presented an empirical study of the word sense disambiguation (WSD). WSD is
as complex as an AI complete problem since it has to dealt with all the complexities of language
in obtaining the semantic structure from the unstructured text. Moreover, the complexity of
WSD is proportional to the granularity of the sense distinctions considered. We have presented
various techniques in Knowledge-based, supervised, and unsupervised approaches. Knowledge
–based approaches appear to be more successful due to highly enriched knowledge sources.
But, it may fail completely if sufficient information is not present in the knowledge sources.
Supervised approaches undoubtedly outperform other approaches. But, they require a large
training corpora for every possibility which demands more human effort and time i.e these may
suffer from knowledge acquisition bottleneck. Unsupervised approaches do not depend on
knowledge sources. So these are most suitable for all word WSD even for languages with low
resources. Following this, we would like to build a comprehensive WSD system for TELUGU
language most spoken Dravidian based language in India.
REFERENCES
[1] LESK.M.1986. Automatic sense disambiguation using machine readable dictionaries: Hoe to tell a
pine cone from an ice cream cone. In Proceedings of the 5th SIGDOC (New York, NY). 24-16
[2] BANERJEE, S. AND PEDERSEN, T. 2003. Extended gloss overlaps as a measure of semantic
relatedness. In Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI,
Acapulco, Mexico). 805–810.
[3] HINDLE, D. AND ROOTH, M. 1993. Structural ambiguity and lexical relations. Computat. Ling. 19,
1, 103–120.
[4] RESNIK, P. S., Ed. 1993. Selection and information: A class-based approach to lexical relationships,
Ph.D. dissertation. University of Pennsylvania, Pennsylvania, Philadelphia, PA.
[5] RESNIK, P. 1997. Selectional preference and sense disambiguation. In Proceedings of the ACL
SIGLEXWorkshop on Tagging Text with Lexical Semantics: Why, What, and How? (Washington,
D.C.). 52–57.
[6] LI, H. AND ABE, N. 1998. Generalizing case frames using a thesaurus and the MDL principle.
Computat.Ling. 24, 2, 217–244.
[7] MCCARTHY, D. AND CARROLL, J. 2003. Disambiguating nouns, verbs and adjectives using
automatically acquired selectional preferences. Computat. Ling. 29, 4, 639–654.
[8] ABNEY, S. AND LIGHT, M. 1999. Hiding a semantic class hierarchy in a Markov model. In
Proceedings of theACL Workshop on Unsupervised Learning in Natural Language Processing
(College Park, MD). 1–8.
[9] FU, K. 1982. Syntactic Pattern Recognition and Applications. Prentice-Hall, Engelwood Cliffs, NJ.
[10]PEDERSEN, T., BANERJEE, S., AND PATWARDHAN, S. 2005. Maximizing semantic relatedness
to perform word sense disambiguation. Res. rep. UMSI 2005/25. University of Minnesota
Supercomputing Institute, Minneapolis, MN.
[11]RADA, R., MILI, H., BICKNELL, E., AND BLETTNER, M. 1989. Development and application of a
metric on semantic nets. IEEE Trans. Syst. Man Cybernet. 19, 1, 17–30.
[12]SUSSNA, M. 1993. Word sense disambiguation for free-text indexing using a massive semantic
network.In Proceedings of the 2nd International Conference on Information and Knowledge Base
Management(Washington D.C.). 67–74.
[13]LEACOCK, C. AND CHODOROW,M. 1998. Combining local context andWordNet similarity for
word sense identification.In WordNet: An electronic Lexical Database, C. Fellbaum, Ed. MIT Press,
Cambridge, MA,265–283.
[14]AGIRRE, E. AND RIGAU, G. 1996.Word sense disambiguation using conceptual density. In
Proceedings of the 16th International Conference on Computational Linguistics (COLING,
Copenhagen, Denmark). 16–22.
[15]RIVEST, R. L. 1987. Learning decision lists. Mach. Learn. 2, 3, 229–246.

41
[16]YAROWSKY, D. 1994. Decision lists for lexical ambiguity resolution: Application to accent
restoration inSpanish and French. In Proceedings of the 32nd Annual Meeting of the Association for
ComputationalLinguistics (Las Cruces, NM). 88–95.
[17]QUINLAN, J. R. 1993. Programs for Machine Learning. Morgan Kaufmann, San Francisco, CA.
[18]QUINLAN, J. R. 1986. Induction of decision trees. Mach. Learn. 1, 1, 81–106.
[19]NG, T. H. 1997. Getting serious about word sense disambiguation. In Proceedings of the ACL
SIGLEXWorkshop on Tagging Text with Lexical Semantics: Why, What, and How? (Washington
D.C.). 1–7.
[20]MCCULLOCH, W. AND PITTS, W. 1943. A logical calculus of the ideas immanent in nervous
activity. Bull.Math. Biophys. 5, 115–133.
[21]COTTRELL, G. W. 1989. A Connectionist Approach to Word Sense Disambiguation. Pitman,
London, U.K.
[22]MOONEY, R. J. 1996. Comparative experiments on disambiguating word senses: An illustration of
the role of bias in machine learning. In Proceedings of the 1996 Conference on Empirical Methods in
Natural Language Processing (EMNLP). 82–91.
[23]TOWELL, G. AND VOORHEES, E. 1998. Disambiguating highly ambiguous words. Computat. Ling.
24, 1, 125–145.
[24]GALE, W. A., CHURCH, K., AND YAROWSKY, D. 1992b. A method for disambiguating word
senses in a corpus.Comput. Human. 26, 415–439.
[25]SCHUTZE, H. 1998. Automatic word sense discrimination. Computat. Ling. 24, 1, 97–124.
[26]SCHUTZE, H. 1992. Dimensions of meaning. In Supercomputing ’92: Proceedings of the 1992
ACM/IEEEConference on Supercomputing. IEEE Computer Society Press, Los Alamitos, CA. 787–
796.
[27]GOLUB, G. H. AND VAN LOAN, C. F. 1989. Matrix Computations. The John Hopkins University
Press,Baltimore, MD.
[28]DEMPSTER, A., LAIRD, N., AND RUBIN, D. 1977. Maximum likelihood from incomplete data via
the EM algorithm. J. Roy. Statist. Soc. 39, 1–38.
[29]PEDERSEN, T. AND BRUCE, R. 1997. Distinguishing word senses in untagged text. In Proceedings
of the1997 Conference on Empirical Methods in Natural Language Processing (EMNLP, Providence,
RI). 197–207.
[30]PURANDARE, A. AND PEDERSEN, T. 2004. Improving word sense discrimination with gloss
augmented feature vectors. In Proceedings of theWorkshop on Lexical Resources for theWeb andWord
Sense Disambiguation(Puebla, Mexico). 123–130.
[31]NIU, C., LI, W., SRIHARI, R., AND LI, H. 2005. Word independent context pair classification model
for wordsense disambiguation. In Proceedings of the 9th Conference on Computational Natural
Language Learning(CoNLL, Ann Arbor, MI).
[32]IDE, N., ERJAVEC, T., AND TUFIS, D. 2001. Automatic sense tagging using parallel corpora. In
Proceedings of the 6th Natural Language Processing Pacific Rim Symposium (Tokyo, Japan). 83–89.
[33]LIN, D. AND PANTEL, P. 2002. Discovering word senses from text. In Proceedings of the 8th ACM
SIGKDDInternational Conference on Knowledge Discovery and Data Mining (Edmonton, Alta.,
Canada). 613–619.
[34]WIDDOWS, D. AND DOROW, B. 2002. A graph model for unsupervised lexical acquisition. In
Proceedings of the 19th International Conference on Computational Linguistics (COLING, Taipei,
Taiwan). 1–7.
[35]VAN DONGEN, S. 2000. Graph Clustering by Flow Simulation, Ph.D. dissertation. University of
Utrecht,Utrecht, The Netherlands.
[36]V´ERONIS, J. 2004. Hyperlex: Lexical cartography for information retrieval. Comput. Speech Lang.
18, 3, 223–252.
[37]NAVIGLI, R. 2009. Word Sense Disambiguation: a Survey, ACM Computing Surveys, Vol. 41, No.2,
ACM Press, pp. 1-69.
[38]Amiat Jain, Sudesh Yadav “Measuring Context-Meaning for Open Class Words in Hindi Language”
Sixth International Conference on Contemporary Computing IEEE 2013. ISBN:978-1-4673-5114-
0,Page(s)173-178

42
Authors
M.Srinivas holds B.Tech in Computer Science and Engineering from JNTU
Hyderabad, M.Tech in Computer Science from JNTU Hyderabad and Pursuing Ph.D in
Computer Science and Engineering from JNTU Hyderabad. His areas of research
interest includes Natural Language Processing, Information Security, Data Mining, and
Data Analytics. At present he is working as Associate Professor, Department of
Computer Science & Engineering, Geethanjali College of Engineering & Technology,
Hyderabad.
Dr.B. Padmaja Rani holds B.E in Electronics and Communication Engineering from
Osmania University, Hyderabad, M.Tech in Computer Science from JNTU
Hyderabad and She received a Doctoral Degree(Ph.D.) in Computer Science from
JNTU Hyderabad. At present she is working as Professor, Department of Computer
Science and Engineering, JNTUH College of Engineering, JNTUH University,
Hyderabad. She is guiding couple of Ph.D Scholars in the area of Information
Retrieval, Natural Language Processing and Information Security. Her area of
research interest includes Information Retrieval, Natural Language Processing,
Information Security, Data Mining and Embedded Systems. To the Credit she published 40 + research
papers in various International/National Conferences and Journals. She is member of various professional
bodies including CSI, IEEE, ISTE.

AN EMPIRICAL STUDY OF WORD SENSE DISAMBIGUATION

More Related Content

What's hot (20)

Viewers also liked (15)

Similar to AN EMPIRICAL STUDY OF WORD SENSE DISAMBIGUATION (20)

Recently uploaded (20)

AN EMPIRICAL STUDY OF WORD SENSE DISAMBIGUATION