Pivot-based Machine Translation between Statistical and Black Box systems

Pivot-based Machine Translation between Statistical and Black Box systems Antonio Toral School of Computing Dublin City University Dublin, Ireland atoral@computing.dcu.ie Abstract This paper presents a novel approach to pivot-based machine translation (MT): while the state-of-the-art uses two statisti- cal systems, this proposal treats the second system as a black box. Our approach ef- fecively provides pivot-based MT to target languages for which no suitable bilingual corpora are available to build statistical systems, as long as any other kind of MT system is available. We experiment with an algorithm that uses two features to find the best translation: the translation score provided by the first system and fluency of the final translation. Despite its simplicity, this approach yields significant improve- ments over the baseline, which translates the source sentences using the two MT sys- tems sequentially. We have experimented with two scenarios, technical documenta- tion in Romance languages and newswire in Slavic languages, obtaining 11.88% and 13.32% relative improvements in terms of BLEU, respectively. 1 Introduction Pivot-based machine translation (MT) refers to the use of an intermediate language, called pivot lan- guage (PL), to translate from the source (SL) to the target language (TL). Unlike typical MT systems, which translate directly from SL to TL, pivot- based systems translate sequentially from SL to PL and then from PL to TL. The main motivation for building pivot-based MT systems is the lack of language resources for a language pair SL–TL, in c⃝ 2012 European Association for Machine Translation. contrast with the availability of such resources for both language pairs SL–PL and PL–TL. Much of the research carried out in pivot-based MT concentrates on a scenario where the transla- tion both from SL to PL and from PL to TL is car- ried out by statistical machine translation systems (SMT). It is also often assumed that the developer has access not only to the output of the systems but also to their internal data structures. Hence, for these methods to work, bilingual corpora for both SL–PL and PL–TL are required in order to train the corresponding SMT systems. Our research concentrates on pivot-based MT for cases where there is no access to the internals of the second system (PL to TL), i.e. we treat it as a black box: only the output translations pro- duced by this system are available. Because of this our approach is applicable to a much broader set of scenarios than the current state-of-the-art; i.e. it can be applied when there is no access to the internals of the second system, which is the case for many online MT systems, or when the second system does not provide the required data (such as n-best lists), which is the case for many rule-based machine translation systems (RBMT). The remainder of this paper is organised as fol- lows. Section 2 presents an overview of the state- of-the-art for pivot-based MT. This is followed by the description of our methodology. Subsequently, we carry out the evaluation and present the results of the …

Content truncated to 500 words