Word Sense Disambiguation by Web mining for word co-occurrence probabilities

Word Sense Disambiguation by Web Mining for Word Co-occurrence Probabilities Peter D. TURNEY Institute for Information Technology National Research Council of Canada Ottawa, Ontario, Canada, K1A 0R6 peter.turney@nrc-cnrc.gc.ca Abstract This paper describes the National Research Coun- cil (NRC) Word Sense Disambiguation (WSD) sys- tem, as applied to the English Lexical Sample (ELS) task in Senseval-3. The NRC system approach- es WSD as a classical supervised machine learn- ing problem, using familiar tools such as the Weka machine learning software and Brill’s rule-based part-of-speech tagger. Head words are represent- ed as feature vectors with several hundred features. Approximately half of the features are syntactic and the other half are semantic. The main novelty in the system is the method for generating the semantic features, based on word co-occurrence probabilities. The probabilities are estimated using the Waterloo MultiText System with a corpus of about one ter- abyte of unlabeled text, collected by a web crawler. 1 Introduction The Senseval-3 English Lexical Sample (ELS) task requires disambiguating 57 words, with an average of roughly 140 training examples and 70 testing examples of each word. Each example is about a paragraph of text, in which the word that is to be dis- ambiguated is marked as the head word. The aver- age head word has around six senses. The training examples are manually classified according to the intended sense of the head word, inferred from the surrounding context. The task is to use the training data and any other relevant information to automat- ically assign classes to the testing examples. This paper presents the National Research Coun- cil (NRC) Word Sense Disambiguation (WSD) system, which generated our four entries for the Senseval-3 ELS task (NRC-Fine, NRC-Fine2, NRC-Coarse, and NRC-Coarse2). Our approach to the ELS task is to treat it as a classical supervised machine learning problem. Each example is repre- sented as a feature vector with several hundred fea- tures. Each of the 57 ambiguous words is represent- ed with a different set of features. Typically, around half of the features are syntactic and the other half are semantic. After the raw examples are converted to feature vectors, the Weka machine learning soft- ware is used to induce a model of the training data and predict the classes of the testing examples (Wit- ten and Frank, 1999). The syntactic features are based on part-of- speech tags, assigned by a rule-based tagger (Brill, 1994). The main innovation of the NRC WSD sys- tem is the method for generating the semantic fea- tures, which are derived from word co-occurrence probabilities. We estimated these probabilities using the Waterloo MultiText System with a corpus of about one terabyte of unlabeled text, collected by a web crawler (Clarke et al., 1995; Clarke and Cor- mack, 2000; Terra and Clarke, 2003). In Section 2, we describe the NRC WSD system. Our experimental results are presented in Section 3 and we conclude in Section 4. 2 System Description This section presents various aspects of the system in roughly the order in …

Content truncated to 500 words