Joining Forces Pays Off: Multilingual Joint Word Sense Disambiguation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 1399–1410, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Joining Forces Pays Off: Multilingual Joint Word Sense Disambiguation Roberto Navigli and Simone Paolo Ponzetto Dipartimento di Informatica Sapienza Universit`a di Roma {navigli,ponzetto}@di.uniroma1.it Abstract We present a multilingual joint approach to Word Sense Disambiguation (WSD). Our method exploits BabelNet, a very large mul- tilingual knowledge base, to perform graph- based WSD across different languages, and brings together empirical evidence from these languages using ensemble methods. The re- sults show that, thanks to complementing wide-coverage multilingual lexical knowledge with robust graph-based algorithms and com- bination methods, we are able to achieve the state of the art in both monolingual and multi- lingual WSD settings. 1 Introduction Nowadays the textual information needed by a user accessing websites for content such as news re- ports, commentaries and encyclopedic knowledge is provided in an increasingly wide range of lan- guages. For example, even though English is still the majority language of the Web, the Chinese and Spanish languages are moving fast to capture their “juicy share”, and more languages are about to join them in the near future. This language explosion clearly forces researchers to focus on the challeng- ing problem of being able to analyze and under- stand text written in any language. However, it also opens up novel perspectives for multilingual Natural Language Processing (NLP) such as, for instance, the development of approaches aimed at “joining forces” and taking advantage of the lexico-semantic knowledge provided in the different languages to improve text understanding. These two aspects are strongly intertwined: on the one hand, enabling language-independent text understanding would al- low for the harvesting of more knowledge in arbi- trary languages, while, on the other hand, bringing together the lexical and semantic information avail- able in different languages would improve the qual- ity of text understanding in arbitrary languages. However, these two goals have hitherto never been achieved, as is attested to by the fact that re- search in a core language understanding task such as Word Sense Disambiguation (Navigli, 2009, WSD) has always been focused mostly on English. His- torically, English became established as the lan- guage used and understood by the scientific com- munity and, consequently, most resources were de- veloped for it, including large-scale computational lexicons like WordNet (Fellbaum, 1998) and sense- tagged corpora like SemCor (Miller et al., 1993). As a result WSD in other languages was hindered by a lack of resources, which in turn led to poor re- sults or low involvement on the part of the research community (Magnini et al., 2004; M`arquez et al., 2004; Orhan et al., 2007; Okumura et al., 2010). Nonetheless, already in the 1990s it had been re- marked that WSD could be improved by means of multilingual information: a recurring idea proposed by several researchers was that plausible transla- tions of a word in context would restrict its pos- …

Content truncated to 500 words