Example-Based Machine Translation of the Basque Language

Example-Based Machine Translation of the Basque Language Nicolas Stroppa, Declan Groves, Andy Way National Centre for Language Technology Dublin City University Dublin 9, Ireland {nstroppa,dgroves,away}@computing.dcu.ie Kepa Sarasola∗ Informatika Fakultatea University of the Basque Country Donostia, Basque Country, Spain kepa.sarasola@ehu.es Abstract Basque is both a minority and a highly in- flected language with free order of sen- tence constituents. Machine Translation of Basque is thus both a real need and a test bed for MT techniques. In this paper, we present a modular Data- Driven MT system which includes differ- ent chunkers as well as chunk aligners which can deal with the free order of sen- tence constituents of Basque. We conducted Basque to English transla- tion experiments, evaluated on a large cor- pus (270, 000 sentence pairs). The ex- perimental results show that our system significantly outperforms state-of-the-art approaches according to several common automatic evaluation metrics. 1 Introduction Machine Translation (MT) has been shown to be useful for minority languages such as Basque. Cur- rently, in the case of Basque, more than ten offi- cial institutions and private companies commonly use translation memories. Moreover, Basque, being a highly inflected language with free order of sen- tence constituents, makes it suitable to be used as a test bed for new MT techniques. As a result, a rule-based MT system has recently been released with the aim of translating texts and ∗ Work carried out while at The National Centre for Lan- guage Technology at Dublin City University in Spring 2006. websites from Spanish to Basque (Alegria et al., 2005).1 Furthermore, the availability of a large Basque-English corpus with more than 270, 000 aligned translation units and with an average of more than 9 words per unit makes it possible to use state- of-the-art corpus-based MT techniques for Basque- English translation. In this paper, we present a new Data-Driven MT system which exploits both EBMT and SMT tech- niques to extract a dataset of aligned chunks. For the extraction of the EBMT data resources, we make use of two different chunking methods. In the case of English, we employ a marker-based chunker based on the Marker Hypothesis (Green, 1979). For Basque, we currently use the dedicated tools devel- oped at the University of the Basque Country while investigating the application of our marker-based chunker to Basque. The chunks are then aligned thanks to a dynamic programming algorithm which is similar to an edit-distance algorithm while allow- ing for block movements (Leusch et al., 2006). This aligner also relies on relationships between chunks, which we compute in several ways. We present a set of Basque to English transla- tion experiments, evaluated on a large corpus con- sisting of software manuals. We show a signif- icant improvement over state-of-the-art SMT sys- tems (Koehn, 2004) according to several common automatic evaluation metrics. In addition, some manual evaluations have been conducted to assess the quality of our extracted aligned resources, and we find that the high quality of our resources con- 1See also http://matxin.sourceforge.net and http://www.opentrad.org. …

Content truncated to 500 words