Processing MWE in Machine Translation Systems

Multi-word expressions include a wide list of categories such as idiomatic expressions, proverbs, collocations, compound words, domain-specific terms, named entities. According to Baldwin and Kim (2010) and (Sag, Baldwin, Bond, Copestake and Flickinger 2002), MWE are lexical items composed of several lexemes which display lexical, syntactic, semantic, pragmatic and/or statistical idiomaticity. Idiomaticity means deviation from the basic properties of the component items: MWEs have various behaviours, more or less predictable from the properties of the individual items. If some categories of MWEs present a high degree of fixedness in word order and their properties (idioms, domain specific terms), other categories such as collocations are more flexible (accepting modifiers, various senses according to context). These expressions are characterised by specific morpho-syntactic properties (preference for some determiners, numbers or genders), strong lexical preference (poser une question but ask a question), specific syntactic behaviour and non-compositional sense (Hausmann, 2004). MWE represent major difficulties for MT systems (Hurskainen, 2008). Most of the MT systems fail to produce correct translations of MWE, while word-to-word translations are not appropriate and the sense is more or less compositional. Whatever the method adopted for translation (statistical, rule-based or example based), producing accurat