Target word selection as proximity in semantic space

Abstract

Lexical selection is a significant problem for wide-coverage machine translation: depending on the context, a given source language word can often be translated into different target language words. In this paper I propose a method for target word selection that assumes the appropriate translation is more similar to the translated context than are the alternatives. Similarity of a word to a context is estimated using a proximity measure in corpus-derived "semantic space". The method is evaluated using an English-Spanish parallel corpus of colloquial dialogue.