Syntactic Features for High Precision Word Sense Disambiguation

Syntactic features for high precision Word Sense Disambiguation David Martínez, Eneko Agirre IxA NLP Group University of the Basque Country Donostia, Spain {jibmaird,eneko}@si.ehu.es Lluís Màrquez TALP Research Center Polytechnical University of Catalonia Barcelona, Spain lluism@lsi.upc.es Abstract This paper explores the contribution of a broad range of syntactic features to WSD: grammatical relations coded as the presence of adjuncts/arguments in isolation or as subcategorization frames, and instantiated grammatical relations between words. We have tested the performance of syntactic features using two different ML algorithms (Decision Lists and AdaBoost) on the Senseval-2 data. Adding syntactic features to a basic set of traditional features improves performance, especially for AdaBoost. In addition, several methods to build arbitrarily high accuracy WSD systems are also tried, showing that syntactic features allow for a precision of 86% and a coverage of 26% or 95% precision and 8% coverage. 1. Introduction Supervised learning has become the most successful paradigm for Word Sense Disambiguation (WSD). This kind of algorithms follows a two-step process: 1. Choosing the representation as a set of features for the context of occurrence of the target word senses. 2. Applying a Machine Learning (ML) algorithm to train on the extracted features and tag the target word in the test examples. Current WSD systems attain high performances for coarse word sense differences (two or three senses) if enough training material is available. In contrast, the performance for finer-grained sense differences (e.g. WordNet senses as used in Senseval 2 (Preiss & Yarowsky, 2001)) is far from application needs. Nevertheless, recent work (Agirre and Martinez, 2001a) shows that it is possible to exploit the precision-coverage trade-off and build a high precision WSD system that tags a limited number of target words with a predefined precision. This paper explores the contribution of a broad set of syntactically motivated features that ranges from the presence of complements and adjuncts, and the detection of subcategorization frames, up to grammatical relations instantiated with specific words. The performance of the syntactic features is measured in isolation and in combination with a basic set of local and topical features (as defined in the literature), and using two ML algorithms: Decision Lists (Dlist) and AdaBoost (Boost). While Dlist does not attempt to combine the features, i.e. it takes the strongest feature only, Boost tries combinations of features and also uses negative evidence, i.e. the absence of features. Additionally, the role of syntactic features in a high-precision WSD system based on the precision-coverage trade-off is also investigated. The paper is structured as follows. Section 2 reviews the features previously used in the literature. Section 3 defines a basic feature set based on the preceding review. Section 4 presents the syntactic features as defined in our work, alongside the parser used. In section 5 the two ML algorithms are presented, as well as the strategies for the precision-coverage trade-off. Section 6 shows the experimental setting and the results. Finally section 7 draws the conclusions and summarizes further work. 2. Previous work. Yarowsky (1994) defined a basic set …

Content truncated to 500 words