Machine Translation by Triangulation: Making Effective Use of Multi-Parallel Corpora

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 728–735, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics Machine Translation by Triangulation: Making Effective Use of Multi-Parallel Corpora Trevor Cohn and Mirella Lapata Human Computer Research Centre, School of Informatics University of Edinburgh {tcohn,mlap}@inf.ed.ac.uk Abstract Current phrase-based SMT systems perform poorly when using small training sets. This is a consequence of unreliable translation es- timates and low coverage over source and target phrases. This paper presents a method which alleviates this problem by exploit- ing multiple translations of the same source phrase. Central to our approach is triangula- tion, the process of translating from a source to a target language via an intermediate third language. This allows the use of a much wider range of parallel corpora for train- ing, and can be combined with a standard phrase-table using conventional smoothing methods. Experimental results demonstrate BLEU improvements for triangulated mod- els over a standard phrase-based system. 1 Introduction Statistical machine translation (Brown et al., 1993) has seen many improvements in recent years, most notably the transition from word- to phrase-based models (Koehn et al., 2003). Modern SMT sys- tems are capable of producing high quality transla- tions when provided with large quantities of training data. With only a small training sample, the trans- lation output is often inferior to the output from us- ing larger corpora because the translation algorithm must rely on more sparse estimates of phrase fre- quencies and must also ‘back-off’ to smaller sized phrases. This often leads to poor choices of target phrases and reduces the coherence of the output. Un- fortunately, parallel corpora are not readily available in large quantities, except for a small subset of the world’s languages (see Resnik and Smith (2003) for discussion), therefore limiting the potential use of current SMT systems. In this paper we provide a means for obtaining more reliable translation frequency estimates from small datasets. We make use of multi-parallel cor- pora (sentence aligned parallel texts over three or more languages). Such corpora are often created by international organisations, the United Nations (UN) being a prime example. They present a chal- lenge for current SMT systems due to their rela- tively moderate size and domain variability (exam- ples of UN texts include policy documents, proceed- ings of meetings, letters, etc.). Our method translates each target phrase, t, first to an intermediate lan- guage, i, and then into the source language, s. We call this two-stage translation process triangulation (Kay, 1997). We present a probabilistic formulation through which we can estimate the desired phrase translation distribution (phrase-table) by marginali- sation, p(s|t) = � i p(s, i|t). As with conventional smoothing methods (Koehn et al., 2003; Foster et al., 2006), triangulation in- creases the robustness of phrase translation esti- mates. In contrast to smoothing, our method allevi- ates data sparseness by exploring additional multi- parallel data rather than adjusting the probabilities of existing data. Importantly, triangulation provides us with separately estimated phrase-tables which could …

Content truncated to 500 words