Sub-sentential exploitation of translation memories

Sub-sentential Exploitation of Translation Memories Michel Simard, Philippe Langlais Laboratoire de recherche appliquée en linguistique informatique (RALI) Département d'Informatique et recherche opérationnelle Université de Montréal C.P. 6128, succursale Centre-ville Montréal (Québec) Canada, H3C 3J7 {SimardM,Felipe}@IRO.UMontreal.CA Abstract Translation memory systems (TMS) are a family of computer tools whose purpose is to facilitate and encourage the re-use of existing translations. By searching a database of past translations, these systems can retrieve the translation of whole segments of text and propose them to the translator for re-use. However, the usefulness of existing TMS’s is limited by the nature of the text segments that that they are able to put in correspondence, generally whole sentences. This article examines the potential of a type of system that is able to recuperate the translation of sub-sentential sequences of words. Keywords Translation memory, translation alignment, translation support tools, example-based machine translation. Introduction Translation memory systems (TMS) are a family of computer tools whose purpose is to facilitate and encourage the re-use of existing translations. Ultimately, within an environment where such a system is used, the same piece of text should never be translated twice (for a given pair of source and target languages). The mechanism that is proposed to attain this goal is to systematically archive the translators' production, as pairs of mutually translated segments. When, within a new text to be translated, a segment of text is encountered that matches some couple in this “translation memory”, its translation can be retrieved and re-used by the translator. A number of vendors now market this sort of TMS, most notably Trados (Translator's WorkBench), IBM (Translation Manager/2), Atril (Déja Vu) and Star-AG (Transit). In all these systems, the couples in the translation memory are normally pairs of sentences. Given a new sentence to be translated, they will usually look for couples whose source-language sentence matches the new sentence in its entirety. This match need not be exact (all the above systems feature some sort of “fuzzy” matching), but obviously, better-matching couples stand a better chance of being re-usable. It is instructive, as Macklovitch and Russell (2000) suggest, to view TM applications in an information retrieval (IR) perspective: when using a TMS, the translator is actually just searching for documents that might help in translating a given sentence. In this case, these “documents” happen to be pairs of mutually translated sentences, and the translator is likely to deem a document relevant if its target-language (TL) segment constitutes an acceptable translation for the source- language (SL) sentence. As for the query, it is constructed automatically by the TMS, from the SL sentence to be translated. The retrieval operation is carried out by matching this query as closely as possible. In this perspective, the default strategy of existing TMS's is an extreme form of high-precision, low-recall search: return only the best matching document, and only if it displays a sufficient resemblance to the source sentence. In IR terms, this is like having an assistant hit the “I'm feeling lucky” button on the …

Content truncated to 500 words