A resource-light phrase scheme for language-portable MT

A resource-light phrase scheme for language-portable MT George Tambouratzis Inst. for Language & Speech Processing / 6 Artemidos & Epidavrou Str., Maroussi,151 25, Athens, Greece giorg_t@ilsp.gr Sokratis Sofianopoulos Inst. for Language & Speech Processing / 6 Artemidos & Epidavrou Str., Maroussi,151 25, Athens, Greece s_sofian@ilsp.gr Marina Vassiliou Inst. for Language & Speech Processing / 6 Artemidos & Epidavrou Str., Maroussi,151 25, Athens, Greece mvas@ilsp.gr Fotini Simistira Inst. for Language & Speech Process- ing / 6 Artemidos & Epidavrou Str., Maroussi,151 25, Athens, Greece fotini@ilsp.gr Nikos Tsimboukakis Inst. for Language & Speech Process- ing / 6 Artemidos & Epidavrou Str., Maroussi,151 25, Athens, Greece ntsimb@ilsp.gr Abstract The present article introduces a phrase- alignment approach that involves the processing of a small bilingual corpus in order to extract suitable structural infor- mation. This is used in the PRESEMT project, whose aim is the quick develop- ment of phrase-based Machine Transla- tion (MT) systems for new language pairs. A main bottleneck of such systems is the need to create compatible parsing schemes in the source and target lan- guages. This bottleneck is overcome by combining two modules, the Phrase aligner module and the Phrasing model generator, both of them being based on pattern recognition principles. 1 Introduction - Summary of the PRE- SEMT MT approach A large proportion of current Machine Transla- tion (MT) systems translate sentences by operat- ing at a sub-sentential level. However, this ne- cessitates either (i) the development of matched segmentations that give similar outputs for the © 2011 European Association for Machine Translation. source and target languages (SL and TL) or (ii) the definition of a mapping between two given segmentations. Both these approaches constrain the MT system to language pairs for which suit- able segmentation schemes exist, which either are or have been made compatible via additional processing. A typical example of such an ap- proach is the METIS-II data-driven MT system (Markantonatou et al., 2006 and Carl et al., 2008), where the parsing tools are used for both the source and the target languages and are ac- cordingly modified. This naturally limits the MT system portability to new languages, due to the need for developing and/or modifying the appro- priate tools for generating the segmentation scheme. The PRESEMT project investigates a novel paradigm, which circumvents this bottleneck, and supports the straightforward development of MT systems for new language pairs, using pat- tern recognition principles. Relying on the use of a large TL monolingual corpus and a small bilin- gual corpus, which typically comprises a few hundred sentences aligned at sentence level, PRESEMT is based on handling sub-sentential segments. It uses a parser only in one language and maps this information to the other language of a given language pair. In other words, given a parser (or more generally a phrasing model) in one of the two languages (either SL or TL), one Mikel L. Forcada, Heidi Depraetere, Vincent Vandeghinste (eds.) Proceedings of the 15th Conference of the European Association for Machine Translation, p. 185192 Leuven, Belgium, …

Content truncated to 500 words