Data-Driven Semantic Analysis for Multilingual WSD and Lexical Selection in Translation

A common way of describing the senses of ambiguous words in multilingual Word Sense Disambiguation (WSD) is by reference to their translation equivalents in another language. The theoretical soundness of the senses induced in this way can, however, be doubted. This type of cross-lingual sense identification has implications for multilingual WSD and MT evaluation as well. In this article, we first present some arguments in favour of a more thorough analysis of the semantic information that may be induced by the equivalents of ambiguous words found in parallel corpora. Then, we present an unsupervised WSD method and a lexical selection method that exploit the results of a data-driven sense induction method. Finally, we show how this automatically acquired information can be exploited for a multilingual WSD and MT evaluation more sensitive to lexical semantics.