Error Driven Word Sense Disambiguation

Error Driven Word Sense Disambiguation Luca Dini and Vittorio Di Tomaso Fr4d4rique Segond CELI Xerox Research Centre Europe {dini, dit omaso}@celi, sns. it Frederique. Segond©xrce. xerox, com Abstract In this paper we describe a method for per- forming word sense disambiguation (WSD). The method relies on unsupervised learning and ex- ploits functional relations among words as pro- duced by a shallow parser. By exploiting an er- ror driven rule learning algorithm (Brill 1997), the system is able to produce rules for WSD, which can be optionally edited by humans in or- der to increase the performance of the system. 1 Introduction Although automatic word sense disambiguation (WSD) remains a much more difficult task than part of speech (POS) disambiguation, resources and automatic systems are starting to appear. Some of these systems are even mature enough to be evaluated. This paper presents an overview of a system for English WSD which will be eval- uated ill the context of the SENSEVAL project 1. We report on performing automatic WSD us- ing a specially-adapted version of Brill's er- ror driven unsupervised learning program (Brill, 1997), originally developed for POS tagging. In our experiment, like ill Resnik (1997), we used both functional and semantic information in or- der to improve the learning capabilities of the system. Indeed, by having access to a syntactic and functional sketch of sentences, and by being able to stipulate which relations are important for sentence meaning, we overcame some of the traditional problems found in continuous bigram models, such as the occurrence of interpolated clauses and passive constructions. Consider, for example, temporal expressions like Tuesday in The stock market Tuesday staged a technical recovery. Such expressions are quite frequent in newspaper text, often appearing near 1 http://www.itri.bton.ac.uk/events/senseval verbs. Without any functional information, the semantic rules produced by the algorithm will stipulate a strong semantic relation between the semantic class of words like Tuesday and the se- mantic class of verbs like stage. On the contrary, if we use information from a shallow parser, we know that Tuesday is an adverbial expression, probably part of the verb phrase, and that the really important relation to learn is the one be- tween the subject and the verb. In the following sections we describe (i) the re- sources we used (Penn Tree Bank, 45 upper level WordNet tags); (ii) the experiment we ran using rule induction techniques on functional relations (functional relation extraction, tag merging, cor- pus preparation and learning); (iii) the evalu- ation we performed on the semantically hand- tagged part of the Brown corpus and, finally, we sketch out the general architecture we are in the process of implementing. 2 The Resources We decided to take advantage of the syntactic structures already contained in the Penn Tree Bank (PTB) (Mitchell et al., 1995) in order to build a large set of functional relation pairs (much as in Resnik (1997)). These relations are then used to learn how to perform semantic dis- ambiguation. To distinguish word meanings we use the …

Content truncated to 500 words