Evaluation of the domain adaptation of MT systems in ACCURAT

Exploiting portability to build an RBMT prototype for a new source language Nora Aranberri, Gorka Labaka, Arantza D´ıaz de Ilarraza and Kepa Sarasola IXA Group University of the Basque Country Manuel Lardizabal 1, 20018 Donostia, Spain {nora.aranberri,gorka.labaka,a.diazdeilarraza,kepa.sarasola}@ehu.eus Abstract This paper presents the work done to port a deep-transfer rule-based machine trans- lation system to translate from a differ- ent source language by maximizing the ex- ploitation of existing resources and by lim- iting the development work. Specifically, we report the changes and effort required in each of the system’s modules to ob- tain an English-Basque translator, ENEUS, starting from the Spanish-Basque Matxin system. We run a human pairwise compar- ison for the new prototype and two statis- tical systems and see that ENEUS is pre- ferred in over 30% of the test sentences. 1 Introduction Building a corpus-based system is undeniably quicker than building a rule-based machine trans- lation (RBMT) system, given the availability of large quantities of parallel text. However, this is often not the case for many language pairs, which makes building a mainstream statistical system suboptimal. Usually, lesser-resourced languages opt for RBMT systems, where language-specific NLP tools and resources are crafted. Heavy investment and long development peri- ods have been attributed to RBMT systems but (Surcin et al., 2013) pointed out that a large part of the systems’ code is reusable. They state that 80% of Systran’s code belongs to the analysis mod- ule, whereas the remaining 20% is equally divided into transfer and generation. Transfer is language- pair specific, but analysis and generation are built c⃝ 2015 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. with information about one language only and they are therefore reusable for systems that use those languages. Rapid development of new language pairs benefits from existing resources but also from modular, stable infrastructures where new pairs can be developed by modifying the linguistic data. An example of RBMT portability attempts for lesser-resourced languages is the Apertium project (Forcada et al., 2011). Apertium is a free/open- source shallow-transfer MT platform. Researchers have been active in porting the system to different language pairs (Peradin et al., 2014; Otte and Ty- ers, 2011). The system specializes in translation between related languages where shallow transfer suffices to produce good quality translations. Shallow parsing is sometimes too limited for dissimilar language pairs. Unrelated languages of- ten require a richer and more flexible deeper trans- fer architecture to tackle differing linguistic fea- tures. Examples are (Gasser, 2012) and Matxin (Mayor et al., 2011). In this work we present an attempt to port the deep-transfer RBMT Matxin1, designed to cope with dissimilar languages. The remaining work is organized as follows: Section 2 gives a brief overview of the architec- ture of the Matxin system. Section 3 describes the work done in each of the system’s modules. Section 4 provides the results of the new Matxin ENEUS prototype’s evaluation. Finally, Section 5 presents the conclusions and future work. 2 …

Content truncated to 500 words