Extracting Glosses to Disambiguate Word Senses

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 627–635, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Extracting Glosses to Disambiguate Word Senses Weisi Duan Carnegie Mellon University Language Technologies Institute 5000 Forbes Ave. Gates Hillman Complex 5407 Pittsburgh, PA 15213 wduan@cs.cmu.edu Alexander Yates Temple University Computer and Information Sciences 1805 N. Broad St. Wachman Hall 303A Philadelphia, PA 19122 yates@temple.edu Abstract Like most natural language disambiguation tasks, word sense disambiguation (WSD) re- quires world knowledge for accurate predic- tions. Several proxies for this knowledge have been investigated, including labeled cor- pora, user-contributed knowledge, and ma- chine readable dictionaries, but each of these proxies requires significant manual effort to create, and they do not cover all of the ambigu- ous terms in a language. We investigate the task of automatically extracting world knowl- edge, in the form of glosses, from an unlabeled corpus. We demonstrate how to use these glosses to automatically label a training cor- pus to build a statistical WSD system that uses no manually-labeled data, with experimental results approaching that of a supervised SVM- based classifier. 1 Introduction For many semantic natural language processing tasks, systems require world knowledge to disam- biguate language utterances. Word sense disam- biguation (WSD) is no exception — systems for WSD require world knowledge to figure out which aspects of a word’s context indicate one sense over another. A fundamental problem for WSD is that the required knowledge is open-ended. That is, for ev- ery ambiguous term, new kinds of information about the world become important, and the knowledge that a system may have acquired for previously-studied ambiguous terms may have little or no impact on the next ambiguous term. Thus open-ended knowledge acquisition is a fundamental obstacle to strong per- formance for this disambiguation task. Researchers have investigated a variety of tech- niques that address this knowledge acquisition bot- tleneck in different ways. Supervised WSD tech- niques, for instance, can learn to associate features in the context of a word with a particular sense of that word. Knowledge-based techniques rely on machine-readable dictionaries or lexical resources like WordNet (Fellbaum, 1998) to provide the nec- essary knowledge. And most recently, systems have used resources like Wikipedia, which contain user-contributed knowledge in the form of sense- disambiguated links, to acquire world knowledge for WSD. Yet each of these approaches is limited by the amount of manual effort that is needed to build the necessary resources, and as a result the techniques are limited to a subset of English words for which the manually-constructed resources are available. In this work we investigate an alternative ap- proach that attacks the problem of knowledge acqui- sition head-on. We use information extraction (IE) techniques to extract glosses, or short textual char- acterizations of the meaning of one sense of a word. In the ideal case, we would extract full logical forms to define word senses, but here we instead focus on a more feasible, but still very …

Content truncated to 500 words