Word Sense Disambiguation Using Sense Examples Automatically Acquired from a Second Language

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 547–554, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Word Sense Disambiguation Using Sense Examples Automatically Acquired from a Second Language Xinglong Wang School of Informatics University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW, UK xwang@inf.ed.ac.uk John Carroll Department of Informatics University of Sussex Falmer, Brighton BN1 9QH, UK johnca@sussex.ac.uk Abstract We present a novel almost-unsupervised approach to the task of Word Sense Dis- ambiguation (WSD). We build sense ex- amples automatically, using large quanti- ties of Chinese text, and English-Chinese and Chinese-English bilingual dictionar- ies, taking advantage of the observation that mappings between words and mean- ings are often different in typologically distant languages. We train a classifier on the sense examples and test it on a gold standard English WSD dataset. The eval- uation gives results that exceed previous state-of-the-art results for comparable sys- tems. We also demonstrate that a little manual effort can improve the quality of sense examples, as measured by WSD ac- curacy. The performance of the classifier on WSD also improves as the number of training sense examples increases. 1 Introduction The results of the recent Senseval-3 competition (Mihalcea et al., 2004) have shown that supervised WSD methods can yield up to 72.9% accuracy1 on words for which manually sense-tagged data are available. However, supervised methods suffer from the so-called knowledge acquisition bottleneck: they need large quantities of high quality annotated data 1This figure refers to the highest accuracy achieved in the Senseval-3 English Lexical Sample task with fine-grained scor- ing. to produce reliable results. Unfortunately, very few sense-tagged corpora are available and manual sense-tagging is extremely costly and labour inten- sive. One way to tackle this problem is trying to automate the sense-tagging process. For example, Agirre et al. (2001) proposed a method for building topic signatures automatically, where a topic signa- ture is a set of words, each associated with some weight, that tend to co-occur with a certain concept. Their system queries an Internet search engine with monosemous synonyms of words that have multiple senses in WordNet (Miller et al., 1990), and then ex- tracts topic signatures by processing text snippets re- turned by the search engine. They trained a classifier on the topic signatures and evaluated it on a WSD task, but the results were disappointing. In recent years, WSD approaches that exploit differences between languages have shown great promise. Several trends are taking place simulta- neously under this multilingual paradigm. A clas- sic one is to acquire sense examples using bilin- gual parallel texts (Gale et al., 1992; Resnik and Yarowsky, 1997; Diab and Resnik, 2002; Ng et al., 2003): given a word-aligned parallel corpus, the dif- ferent translations in a target language serve as the “sense tags” of an ambiguous word in the source language. For example, Ng et al. (2003) acquired sense examples using English-Chinese parallel cor- pora, which were manually or automatically aligned at sentence level and then …

Content truncated to 500 words