Article selection using probabilistic sense disambiguation

MT Summit VII Sept.1999 Article Selection Using Probabilistic Sense Disambiguation Lee Hian-Beng DSO National Laboratories 20 Science Park Drive, Singapore 118230 Abstract A probabilistic method is used for word sense disambiguation where the features taken are the surrounding six words. As their surface forms are used, no syntac- tic or semantic analysis is required. De- spite its simplicity, this method is able to disambiguate the noun interest accurately. Using the common data set of (Bruce & Wiebe 94), we have obtained an average accuracy of 86.6% compared with their re- ported figure of 78%. This portable tech- nique can be applied to the task of English article selection. This problem arises from machine translation of any source language without article to English. Using texts from the Wall Street Journal, we achieved an overall accuracy of 83.1% for the 1,500 most commonly used head nouns. 1 Introduction Word sense disambiguation is an important problem in natural language processing, which has been ad- dressed in a variety of ways. A popular strategy is the knowledge-based approach, in which linguistic knowl- edge is used to achieve disambiguation. For exam- ple (Agirre &: Rigau 95) have made use of the Word- Net noun hierarchy to select the word sense which shares the same WordNet noun subtree as most of the surrounding nouns. (Véronis & Ide 95) took advan- tage of machine-readable dictionaries instead, identi- fying similarities between words through definitions in the dictionary. The thesaurus has also been used by Yarowsky 92) and (Okumura & Honda 94), whose semantic categories serve as classes for disambigua- tion. (Wilks & Stevenson 98) used knowledge such as dictionary definitions, pragmatic codes and selectional restrictions for word sense disambiguation. However, these approaches all require a WordNet-like lexical knowledge source or a machine-readable dictionary, which often take years to construct. Relying on such resources also makes the extension of these methods to other languages difficult. Recently, much attention has been focused on the statistical approach to solving the problem. In par- ticular, (Bruce & Wiebe 94) have used a probabilistic classifier based on keywords, as well as syntactic in- formation such as morphology and parts of speech. (Yarowsky 92) has modeled the sense-disambiguation problem using the naive Bayesian classifier, where the features used are words in the surrounding 100-word window. Instead of this set of features, (Leacock et al. 93) have chosen to use two-sentence contexts: the sentence containing the word to be disambiguated and the preceding sentence. Although humans appear to require only a few words for sense resolution (Choueka & Lusignan 85), large window sizes have been found to be necessary in these studies presumably because so much information, such as word-order and syntax, has been thrown away. (Mooney 96) has also studied the Bayesian classifier and compared it with other meth- ods, although in the preprocessing he reduces words to stems and removes stopwords. In this paper, we present a probabilistic disam- biguation algorithm which does not require any syn- tactic or semantic information (Teo et …

Content truncated to 500 words