Word sense disambiguation criteria: a systematic study

Word sense disambiguation criteria: a systematic study Laurent AUDIBERT DELIC, Université de Provence 29 Av. Robert SCHUMAN – 13621 Aix-en-Provence Cedex 1, FRANCE laurent.audibert@up.univ-aix.fr Abstract This article describes the results of a systematic in- depth study of the criteria used for word sense disambiguation. Our study is based on 60 target words: 20 nouns, 20 adjectives and 20 verbs. Our results are not always in line with some practices in the field. For example, we show that omitting non- content words decreases performance and that bigrams yield better results than unigrams. 1 Introduction The task of word sense disambiguation (WSD) is to identify the correct sense of a word in context. WSD is usually performed by matching information from the context in which the word occurs with disambiguation knowledge source. Our approach uses supervised machine-learning techniques to automatically acquire such disambiguation knowledge from sense-tagged corpora. At present, this type of approach is widely used and seems to yield the best results (Kilgarriff, Rosenzweig, 2000; Ng, 1997b). Information conveyed by context words (morphological form) is enriched with further annotations: part-of-speech tag, lemma, etc. Each individual piece of information is called a feature. A good feature should capture an important source of knowledge that is critical in determining the sense of the word to be disambiguated. The choice of the appropriate set of features is an important issue for WSD (Bruce, Wiebe, Perdersen, 1996; Ng, Zelle, 1997; Pedersen, 2001). Thus, this paper describes the results of a systematic and in-depth study of homogenous criteria (i.e. set of features) that can be used for WSD. 2 Methodology 2.1 Corpus The corpus we worked on is composed of different types of texts and comprises 6 468 522 words. It was put together within the framework of the SyntSem project that aims at producing a French corpus which is morphologically and syntactically tagged, lemmatised and that comprises a light syntactical tagging as well as a lexical tagging of 60 target words selected for their strongly polysemous nature (Véronis, 1998)1. These 60 target words are evenly divided into 20 nouns, 20 adjectives and 20 verbs, having a total of 53 796 occurrences in the corpus. The inadequacy of standard dictionaries (Véronis, 2001) and computational lexicons (Palmer, 1998) for natural language processing is presently one of the major difficulties encountered in word sense disambiguation. For instance, by using these dictionaries, the inter-annotator agreement may sometimes reach only 57% (Ng, Lee, 1996) or may simply be equivalent to random sense allocation (Véronis, 1998). To overcome this weakness, a dictionary more specific to natural language processing is being developed in our team (Reymond, 2002). It has been used to tag the occurrences of the 60 target words of the SyntSem corpus. Table 8 in the appendix gives quantitative information for each target word. The number of senses per word may be very large for it includes idioms and phrasal verbs such as: « mettre sur pied », « mettre à pied », « pied de nez », etc. A …

Content truncated to 500 words