Semantic Tagging Using WordNet Examples

Semantic Tagging Using W ordN et Examples Sherwood Haynes Illinois Institute of Technology Department of Computer Science Chicago, Illinois, 60616 USA skhii@mindspring.com Abstract This paper describes IITl, IIT2, and IIT3, three versions of a semantic tagging system basing its sense discriminations on WordNet examples. The system uses WordNet relations aggressively, both in identifying examples of words with similar lexical constraints and matching those examples to the context. 1 Introduction The ability of natural language understanding systems to determine the meaning of words in context has long been suggested as a necessary precursor to a deep understanding of the context (Ide and Veronis, 1998; Wilks, 1988). Competitions such as SENSEV AL (Kilgarriff anp Palmer, 2000) and SENSEV AL-2 (SENSEV AL-2, 2001) model the determination of word meaning as a choice of one or more items from a fixed sense inventory, comparing a gold standard based on human judgment to the performance of computational word sense disambiguation systems. Statistically based systems that train on tagged data have regularly performed best on these tasks (Kilgarriff and Rosenzweig, 2000). The difficulty with these supervised systems is their insatiable need for reliable annotated data, frequently called the "data acquisition bottleneck." The systems described here avoid the data acquisition bottleneck by using only a sense repository, or more specifically the examples and relationships contained in the sense repository. WordNet version 1.7 (Miller 1990; Fellbaum 1998; WordNet, 2001) was chosen as the sense repository for the English Lexical Sample task (where systems disambiguate a single word or collocation in context) and the English All Word 79 task (where systems disambiguate all content words) of the SENSEV AL-2 competition. WordNet defmes a word sense (or synset) as a collection of words that can express the sense, a definition of the sense (called a gloss), zero or more examples of the use of the word sense, and a set of tuples that defme relations between synsets or synset words. 2 General Approach This paper describes three systems that were entered in SENSEVAL-2 competition, IITl, IIT2, and IIT3. liT 1 and IIT2 were entered in both the English All Word task and the English Lexical Sample task. IIT3 was entered in the English All Word task only. All three systems use the same unsupervised approach to determine the sense of a target word: 1. for each syntactically plausible sense, fmd the set of WordNet examples that appear in that synset or a related synset. 2. for each example, compare the example to the context, scoring the quality of the match. 3. choose the sense whose synset is responsible for the inclusion of the highest scoring example. Hereafter, target words identify the words to be disambiguated (so identified by the SENSEV AL-2 task). The context identifies the text surrounding and including a target word. 2.1 Collecting Examples of a Sense The systems first collect a set of example sentences and phrases from WordNet for each synset matching a target word (or its canonical or collocational form). The set includes examples from …

Content truncated to 500 words