Use of Rich Linguistic Information to Translate Prepositions and Grammar Cases to Basque

Proceedings of the 13th Annual Conference of the EAMT, pages 58–65, Barcelona, May 2009 Use of Rich Linguistic Information to Translate Prepositions and Grammatical Cases to Basque Eneko Agirre, Aitziber Atutxa, Gorka Labaka, Mikel Lersundi, Aingeru Mayor, Kepa Sarasola IXA Group. University of the Basque Country aingeru@ehu.es Abstract This paper presents three successful tech- niques to translate prepositions heading verbal complements by means of rich lin- guistic information, in the context of a rule-based Machine Translation system for an agglutinative language with scarce re- sources. This information comes in the form of lexicalized syntactic dependency triples, verb subcategorization and manu- ally coded selection rules based on lex- ical, syntactic and semantic information. The first two resources have been auto- matically extracted from monolingual cor- pora. The results obtained using a new evaluation methodology show that all pro- posed techniques improve precision over the baselines, including a translation dic- tionary compiled from an aligned corpus, and a state-of-the-art statistical Machine Translation system. The results also show that linguistic information in all three tech- niques are complementary, and that a com- bination of them obtains the best F-score results overall. 1 Introduction Since the first Machine Translation (MT) systems up to today’s, performing well the translation of the prepositions is relevant for any MT system; Japkowicz and Wiebe (1991) claimed that doing it correctly is difficult because prepositions can- not be translated in a systematic or coherent way. Koehn (2003) remarked the importance of the cor- rect translation of prepositions and he also reported that the main reason for noun phrase (NP) and c⃝ 2009 European Association for Machine Translation. prepositional phrase (PP) mistranslations consists of choosing wrong leading preposition. Translation of prepositions is even more com- plex when the verb phrase and prepositional phrase structures differ widely in the languages involved in translation (Naskar and Bandyopad- hyayn, 2006). This is what happens when trans- lating from Spanish or English into Basque. This paper explores the problem of translat- ing prepositions heading verbal complements into target language equivalents. Although we focus on Spanish to Basque translation, the evaluation methodology and techniques can be applied to other language pairs. In Basque syntactic func- tions like subject, object and indirect objects are marked by case-suffixes. In this work postposi- tions and grammatical cases have been homoge- neously treated, therefore it covers not only the translation of Spanish prepositions, but also how to choose the correct grammatical case correspond- ing to Spanish subjects, objects and indirect ob- jects. Note that in most of the cases Spanish sub- jects and objects are not marked by any surface word or special case marking. Thus, besides the Spanish prepositions, we also explore the transla- tion of the zero preposition corresponding to the grammatical cases of subject and object. Given an existing open-source rule-based ma- chine translation (RBMT) system called Matxin (Mayor, 2007; Alegria et al., 2007), we propose and evaluate three different techniques for trans- lating Spanish prepositions and syntactic functions into Basque. These techniques use rich linguistic …

Content truncated to 500 words