Deep Linguistic Multilingual Translation and Bilingual Dictionaries

Proceedings of the Fourth Workshop on Statistical Machine Translation , pages 90–94, Athens, Greece, 30 March – 31 March 2009. c⃝2009 Association for Computational Linguistics Deep Linguistic Multilingual Translation and Bilingual Dictionaries Eric Wehrli, Luka Nerima & Yves Scherrer LATL-Department of Linguistics University of Geneva fEric.Wehrli, Luka.Nerima, Yves.Scherrerg@unige.ch Abstract This paper describes the MulTra project, aiming at the development of an efficient multilingual translation technology based on an abstract and generic linguistic model as well as on object-oriented software de- sign. In particular, we will address the is- sue of the rapid growth both of the trans- fer modules and of the bilingual databases. For the latter, we will show that a signifi- cant part of bilingual lexical databases can be derived automatically through transitiv- ity, with corpus validation. 1 Introduction The goal of the MulTra project is to develop a grammar-based translation model capable of han- dling not just a couple of languages, but poten- tially a large number of languages. This is not an original goal, but as 50 years of work and in- vestment have shown, the task is by no means an easy one, and although SMT has shown fast and impressive results towards it (e.g. EuroMatrix), we believe that a (principled) grammar-based ap- proach is worth developing, taking advantage of the remarkable similarities displayed by languages at an abstract level of representation. In the first phase of this project (2007-2009), our work has focused on French, English, German, Italian and Spanish, with preliminary steps towards Greek, Romanian, Russian and Japanese. To evaluate the quality of the (still under devel- opment) system, we decided to join the WMT09 translation evaluation with prototypes for the fol- lowing language pairs: English to French, French to English and German to English. In this short paper, we will first give a rough description of the MulTra system architecture and then turn to the difficult issue of the bilingual dictionaries. The MulTra project relies to a large extent on abstract linguistics, inspired from recent work in generative grammar (Chomsky, 1995, Culicover & Jackendoff, 2005, Bresnan, 2001). The grammar formalism developed for this project is both rich enough to express the structural diversity of all the languages taken into account, and abstract enough to capture the generalizations hidden behind ob- vious surface diversity. At the software level, an object-oriented design has been used, similar in many ways to the one adopted for the multilingual parser (cf. Wehrli, 2007). The rapid growth of the number of transfer modules has often been viewed as a major flaw of the transfer model when applied to multilingual translation (cf. Arnold, 2000, Kay, 1997). This ar- gument, which relies on the fact that the number of transfer modules and of the corresponding bilin- gual dictionaries increases as a quadratic function of the number of languages, is considerably weak- ened if one can show that transfer modules can be made relatively simple and light (cf. section 2), compared to the analysis and generation modules (whose numbers are a …

Content truncated to 500 words