Relevance of Different Segmentation Options on Spanish-Basque SMT

Proceedings of the 13th Annual Conference of the EAMT, pages 74–80, Barcelona, May 2009 Relevance of Different Segmentation Options on Spanish-Basque SMT Arantza D´ıaz de Ilarraza, Gorka Labaka and Kepa Sarasola Euskal Herriko Univertsitatea/Universidad del Pa´ıs Vasco jipdisaa@ehu.es, gorka.labaka@ehu.es, jipsagak@ehu.es Abstract Segmentation is widely used in adapting Statistical Machine Translation to highly inflected languages as Basque. The way this segmentation is carried out impacts on the quality of the translation. In or- der to look for the most adequate segmen- tation for a Spanish-Basque system, we have tried different segmentation options and analyzed their effects on the transla- tion quality. Although all segmentation options used in this work are based on the same morpho- logical analysis, translation quality varies significantly depending on the segmen- tation criteria used. Most of the seg- mentation options outperform the base- line according to all metrics, except the one which splits words according the mor- pheme boundaries. From here we can con- clude the importance of the development of the segmentation criteria in SMT. 1 Introduction In this paper we present the work done for adapting a baseline SMT system to carry out the translation into a morphologically-rich agglutinative language such as Basque. In translation from Spanish to Basque, some Spanish words, such as prepositions or articles, correspond to Basque suffixes, and, in case of ellipsis, more than one of those suffixes can be added to the same word. In this way, based on the Basque lemma ’etxe’ /house/ we can generate ’etxeko’ /of the house/, ’etxekoa’ /the one of the house/, ’etxekoarengana’ /towards the one of the house/ and so on. c⃝ 2009 European Association for Machine Translation. Besides, Basque is a low-density language and there are few corpora available comparing to other languages more widely used as Spanish, English, or Chinese. For instance, the parallel corpus avail- able for this work is 1M word for Basque (1.2M words for Spanish), much smaller than the corpora usually used on public evaluation campaigns such as NIST. In order to deal with the problems presented above, we have split up Basque words into the lemma and some tags which represent the mor- phological information expressed on the inflection. Dividing Basque words in this way, we expect to reduce the sparseness produced by the agglutina- tive being of Basque and the small amount of train- ing data. Anyway, there are several options to define Basque segmentation. For example, considering all the suffixes all together as a unique segment, considering each suffix as a different segment, or considering any other of their intermediate com- binations. In order to define the most adequate segmentation for our Spanish-Basque system, we have tried some of those segmentation options and have measured their impact on the translation qual- ity. The remainder of this paper is organized as fol- lows. In Section 2, we present a brief analysis of previous works adapting SMT to highly inflected languages. In Section 3, we describe the systems developed for this paper (the baseline and the mor- …

Content truncated to 500 words