Abstract
We introduce a method for leaning to assign word senses to bilingual translation pairs. In our approach, this problem is transformed into a problem on how to navigate through a sense network (e.g., WordNet) aimed at relating the features of translations to the sense nodes in the network. The method involves automatically constructing classification models for each branched nodes in the sense network and learning to reject less probable sense categories for the translations based on the translation characteristics of semantically related word groups (e.g., words in a lexical category). At run-time, given translations are expanded with their synonyms and the sense ambiguity is resolved according to the trained classification models. Evaluation shows that the method significantly outperforms the strong baseline of assigning most frequent sense to the translation pairs. Our method effectively determines adequate word senses for given word-translation pairs, suggesting the possibility of using our methods as computer-assisted tool for lexicography or of using our method to assist machine translation systems in word selection.