We modify the Lesk algorithm (Lesk, 1986) in several ways to create our baseline algorithm. The Lesk algorithm relies on glosses found in tradi­tional dictionaries which often do not have enough words for the algorithm to work well. We choose the lexical database WordNet, to take advantage of the highly inter-connected set of relations among different words that WordNet offers, and Extended WordNet to capitalize on its (POS and sense) tagged glosses.

The Lesk algorithm takes a local approach for sense disambiguation. The disambiguation of the various words in a sentence is a series of inde­pendent problems and has no effect on each other. We propose a global approach where all the words (we mean by word, an open-class lemma) in the context window are simultaneously disambiguated in a bid to get the best combination of senses for all the words in the window instead of only the target word. The process can be thought of as sense disambiguation of the whole context, instead of a word.

The Lesk algorithm disambiguates words in short phrases. But, the basic unit of disambiguation of our algorithm is the entire sentence under consid­eration. We later modify the context to include the previous and next sentence.

Another major change is that the dictionary definition or gloss of each of its senses is com­pared to the glosses of every other word in the con­text by the Lesk algorithm. But in the present work, the words themselves are compared with the glosses of every other word in the context.

While Lesk's algorithm restricts its comparisons to the dictionary meanings of the words being disam-biguated, our choice of dictionary allows us to also compare the meanings (i.e., glosses) of the words, as well as the words that are related to them through various relationships defined in WordNet. For each POS we choose a relation if links of its kind form at least 5% of the total number of links for that part of speech, with two exceptions. We use the attribute relation although there are not many links of its kind. But this relation links adjec­tives, which are not well developed in WordNet, to nouns which have a lot of data about them. This potential to tap into the rich noun data prompted us to use this relation. Another exception is the an-tonymy relationship. Although there are sufficient antonymy links for adjectives and adverbs, we have not utilized these relations.

Table 1. WordNet relations chosen for the disam­biguation algorithm

The gloss bag is constructed for every sense of every word in the sentence. The gloss-bag is con­structed from the POS and sense tagged glosses of synsets, obtained from the Extended WordNet. For any synset, the words forming the synset and the gloss definition contribute to the gloss-bag. The non-content words are left out. Example sentences do not contribute to the gloss bag since they are not (POS and sense) tagged. Each word along with its POS and sense-tag are stored in the gloss bag. For words with different POS, different relations are taken into account (according to Table 1) for build­ing the corresponding gloss-bag.

This gloss-bag creation process can be per­formed offline or online. It can be performed dy­namically on a as-when-needed basis. Or, gloss-bags can be created for all WordNet entries only once and stored in a data file in prior. The issue is time versus space.

Once, this gloss-bag creation process is over, the comparison process starts. Each word (say Wi) in the context is compared with each word in the gloss-bag for every sense (say Sk) of every other word (say Wj) in the context. If a match is found, they are checked further for part-of-speech match. If the words match in part-of-speech as well, a score is assigned to both the words: the word being matched (Wi) and the word whose gloss-bag con­tains the match (Wj). This matching event indicates mutual confidence towards each other, so both words are rewarded for this event. Two two-dimensional (one for word index and the other for sense index) vectors are maintained: sense_vote for the word in context, and sense_score for the word in gloss-bag. Say, for example, the context word (Wi # noun) matches with gloss word (Wn # noun # m) (i.e., Wi = Wn) in the gloss bag for kth sense of Wj. Then, a score of 1/(gloss bag size of (WJk)) is assigned to both sensevote [i][m] and sense score[j][k]. Scores are normalized before assigning because of huge discrepancy in gloss-bag sizes. This process continues until each context word is matched against all gloss-bag words for each sense of every other context words.

Once all the comparisons have been made, we add sense_vote value with the sense_score linearly value for each sense of every word to arrive at the combination score for this word-sense pair.

The algorithm assigns a word the nth sense for which the corresponding sense_vote and sense_score produces the maximum sum, and it does not assign a word any sense when the corre­sponding sense_vote and sense_score values are 0, even if the word has only one sense. In the event of a tie, we choose the one that is more frequent, as specified by WordNet.

Assuming that there are N words in the window of context (i.e. the sentence), and that, on an aver­age there are S senses per word, and G number of gloss words in each gloss bag per sense, N * S gloss bags need to be constructed, giving rise to a total of N * S * G gloss words. Now these many gloss words are compared against each of the N context words. Thus, N * S * G pairs of word comparisons need to be performed. Both, S and G vary heavily.

The poor performance of the algorithm perhaps suggests that sentential context is not enough for this algorithm to work. So we went for a larger context: a context window containing the current sentence under consideration (target sentence), its preceding sentence and the succeeding sentence. This increment in context size indeed performed better than the baseline algorithm.

When constructing the gloss-bags for a word-sense pair, some words may appear in more than one gloss (by gloss we mean to say synonyms as well as gloss). So, we added another parameter with every (word#pos#sense) in a gloss bag: noc - the number of occurrence of this (word#pos#sense) combination in this gloss-bag.

And, in case of a match of context word (say Wi) with a gloss-bag word (of say kth sense of word Wj), we scored the words in four ways to see if this phenomenon has any effect on the sense disam­biguation process. Say, for example, the context word (Wi- # noun) matches with gloss word (Wn # noun # m # noc) in the gloss bag for kth sense of Wj (i.e., the particular word appears noc times in the said gloss-bag) and the gloss bag size is gbs. Then, we reward Wi and Wj for this event in four ways given below.