Improvements in Analogical Learning: Application to Translating Multi-Terms of the Medical Domain

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 487–495, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Improvements in Analogical Learning: Application to Translating multi-Terms of the Medical Domain Philippe Langlais DIRO Univ. of Montreal, Canada felipe@iro.umontreal.ca Franc¸ois Yvon and Pierre Zweigenbaum LIMSI-CNRS Univ. Paris-Sud XI, France {yvon,pz}@limsi.fr Abstract Handling terminology is an important matter in a translation workflow. However, current Machine Translation (MT) sys- tems do not yet propose anything proactive upon tools which assist in managing termi- nological databases. In this work, we in- vestigate several enhancements to analog- ical learning and test our implementation on translating medical terms. We show that the analogical engine works equally well when translating from and into a mor- phologically rich language, or when deal- ing with language pairs written in differ- ent scripts. Combining it with a phrase- based statistical engine leads to significant improvements. 1 Introduction If machine translation is to meet commercial needs, it must offer a sensible approach to trans- lating terms. Currently, MT systems offer at best database management tools which allow a human (typically a translator, a terminologist or even the vendor of the system) to specify bilingual ter- minological entries. More advanced tools are meant to identify inconsistencies in terminological translations and might prove useful in controlled- language situations (Itagaki et al., 2007). One approach to translate terms consists in us- ing a domain-specific parallel corpus with stan- dard alignment techniques (Brown et al., 1993) to mine new translations. Massive amounts of par- allel data are certainly available in several pairs of languages for domains such as parliament de- bates or the like. However, having at our disposal a domain-specific (e.g. computer science) bitext with an adequate coverage is another issue. One might argue that domain-specific comparable (or perhaps unrelated) corpora are easier to acquire, in which case context-vector techniques (Rapp, 1995; Fung and McKeown, 1997) can be used to identify the translation of terms. We certainly agree with that point of view to a certain extent, but as discussed by Morin et al. (2007), for many specific domains and pairs of languages, such re- sources simply do not exist. Furthermore, the task of translation identification is more difficult and error-prone. Analogical learning has recently regained some interest in the NLP community. Lepage and De- noual (2005) proposed a machine translation sys- tem entirely based on the concept of formal anal- ogy, that is, analogy on forms. Stroppa and Yvon (2005) applied analogical learning to sev- eral morphological tasks also involving analogies on words. Langlais and Patry (2007) applied it to the task of translating unknown words in several European languages, an idea investigated as well by Denoual (2007) for a Japanese to English trans- lation task. In this study, we improve the state-of-the-art of analogical learning by (i) proposing a simple yet effective implementation of an analogical solver; (ii) proposing an efficient solution to the search is- sue embedded in analogical learning, (iii) investi- …

Content truncated to 500 words