Cohesion can be defined as the way certain words or grammatical features of a sentence can connect it to its predecessors (and successors) in a text. (Halliday and Hasan, 1976) defined cohesion as "the set of possibilities that exist in the language for making text hang together". Cohesion occurs where the interpretation of some element in the discourse is dependent on that of another. For ex­ample, an understanding of the reference of a pro­noun (i.e.: he, she, it, etc.) requires to look back to something that has been said before. Through this cohesion relation, two text clauses are linked to­gether.

Cohesion is achieved through the use in the text of semantically related terms, reference, ellipse and conjunctions (Barzilay and Elhadad, 1997). Among the different cohesion-building devices, the most easily identifiable and the most frequent type is lexical cohesion. Lexical cohesion is created by using semantically related words (repetitions, synonyms, hypernyms, hyponyms, meronyms and holonyms, glosses, etc.)

Our technique used WordNet (Miller, 1990) as the knowledge source to find the semantic relations among the words in a text. We assign weights to the semantic relations. The technique can be de­composed into two steps: (1) building a representa­tion of all possible senses of the words and (2) dis-ambiguating the words based on the highest score. The remainder of this paper is organized as fol­lows. In the next section, we review previous work. In Section 3, we define the semantic relations and their weights. Section 4 presents our two step pro­cedure for WSD. We conclude with the evaluation.

Lexical Chaining is the process of connecting se-mantically related words, creating a set of chains that represent different threads of cohesion through the text (Galley and McKeown, 2003). This inter­mediate representation of text has been used in many natural language processing applications, including automatic summarization (Barzilay and Elhadad, 1997; Silber and McCoy, 2003), informa­tion retrieval (Al-Halimi and Kazman, 1998), and intelligent spell checking (Hirst and St-Onge, 1998).

Morris and Hirst (1991) at first proposed a man­ual method for computing lexical chains and first computational model of lexical chains was intro­duced by Hirst and St-Onge (1997). This linear­time algorithm, however, suffers from inaccurate WSD, since their greedy strategy immediately dis-ambiguates a word as it is first encountered. Later research (Barzilay and Elhadad, 1997) significantly alleviated this problem at the cost of a worse run­ning time (quadratic); computational inefficiency is due to their processing of many possible combina­tions of word senses in the text in order to decide which assignment is the most likely. Silber and McCoy (2003) presented an efficient linear-time algorithm to compute lexical chains, which models Barzilay's approach, but nonetheless has inaccura­cies in WSD.

More recently, Galley and McKeown (2003) suggested an efficient chaining method that sepa­rated WSD from the actual chaining. It performs the WSD before the construction of the chains. They showed that it could achieve more accuracy than the earlier ones. Our method follows the simi­lar technique with some new semantic relations (i.e.: gloss, holonym, meronym).

Semantic Relations

We used WordNet2.1 (Miller, 1990) and extended WordNet (Moldovan and Mihalcea, 2001) as our knowledge source to find the semantic relations among the words in a context. We assigned a weight to each semantic relation. The relations and their scores are summarized in the table 1.

System Overview

The global architecture of our system is shown in Figure 1. Each of the modules of the system is de­scribed below.

Context Processing

Context-processing involves preprocessing the con­texts using several tools. We have used the follow­ing tools:

Extracting the main text: This module extracts the context of the target word from the source xml document removing the unnecessary tags and makes the context ready for further processing.

Sentence Splitting, Text Stemming and Chunking: This module splits the context into sen­tences, then stems out the words and chunks those. We used OAK systems (Sekine, 2002) for this purpose.

http:// wordnet.princeton.edu/

Candidate Words Extraction: This module ex­tracts the candidate words (for task 7: noun, verb, adjective and adverb) from the chunked text.

In SemEval-2007, we participated in Task 7: "Coarse-grained English all-words". The evalua­tion of our system is given below:

In this paper, we presented briefly our WSD sys­tem in the context of SemEval 2007 Task 7. Along with normal WordNet relations, our method also included additional relations such as repetition and gloss using semantically enhanced tool, extended WordNet. After disambiguation, the intermediate representation (disambiguation graph) can be used to build the lexical chains which in tern can be used as an intermediate representation for other NLP applications such as text summarization, question answering, text clustering. This method (summing edge weights in selecting the right sense) of WSD before constructing the chain (Gallery and McKe-own, 2003) outperforms the earlier methods of Barzilay and Elhadad (1997) and Silber and McCoy (2003) but this method is highly dependent on the lexical cohesion among words in a context. So the length of context is an important factor for our system to achieve good performance. For the task the context given for a tagged word was not so large to capture the semantic relations among words. This may be the one of the reasons for which our system could not achieve one of the best results.

Chali, Y. and Kolla, M. 2004. Summarization techniques at DUC 2004. In Proceedings of the Document Understanding Conference, pages 105 -111, Boston. NIST.

Galley, M. and McKeown, K. 2003. Improving Word Sense Disambiguation in Lexical Chain­ing. In Proceedings of the 18th International Joint Conference on Artificial Intelligence (IJCAI'03), pages 1486-1488, Acapulco, Mex­ico.

Halliday M. and Hasan R. 1976. Cohesion in Eng­lish. Longman, London.

Harabagiu S. and Moldovan D. 1998. WordNet: An Electronic Lexical Database, chapter Knowl­edge Processing on an Extended WordNet. MIT press.

Hirst G. and St-Onge D. 1997. Lexical Chains as representation of context for the detection and correction of malapropisms. In Christiane Fell­baum, editor, WordNet: An Electronic Lexical Database and Some of its Applications. MIT Press, pages 305-332.

Morris J. and Hirst. G. 1991, Lexical Cohesion Computed by Thesaural Relations as an Indica­tor of the Structure of Text .Computational Lin­guistics, 17(1):21-48.

Silber H.G. and McCoy K.F. 2002. Efficiently Com­puted Lexical Chains As an Intermediate Representa­tion for Automatic Text Summarization. Computa­tional Linguistics, 28(4):487-496.