Extracting collocations and their translations from parallel corpora

Identifying collocations in a text (e.g., break record) and correctly translating them (battre record vs. *casser record) represent key issues in machine translation, notably because of their prevalence in language and their syntactic flexibility. This article describes a method for discovering translation equivalents for collocations from parallel corpora, aimed at increasing the lexical coverage of a machine translation system. The method is based on a “deep” syntactic approach, in which collocations and candidate translations are identified from sentence-aligned text with the help of a multilingual parser. The article also introduces the tools on which this method relies. It focuses in particular on the efforts made to account for structural divergences between languages and to improve the method’s performance in terms of coverage. MOTS-CLÉS : collocations, équivalents de traduction, analyse syntaxique, alignement de texte.