Optimal Bilingual Data for French-English PB-SMT

Proceedings of the 13th Annual Conference of the EAMT, pages 96–103, Barcelona, May 2009 Optimal Bilingual Data for French–English PB-SMT Sylwia Ozdowska and Andy Way National Centre for Language Technology Dublin City University Glasnevin, Dublin 9, Ireland {sozdowska,away}@computing.dcu.ie Abstract We investigate the impact of the original source language (SL) on French–English PB-SMT. We train four configurations of a state-of-the-art PB-SMT system based on French–English parallel corpora which differ in terms of the original SL, and con- duct experiments in both translation direc- tions. We see that data containing orig- inal French and English translated from French is optimal when building a system translating from French into English. Con- versely, using data comprising exclusively French and English translated from several other languages is suboptimal regardless of the translation direction. Accordingly, the clamour for more data needs to be tem- pered somewhat; unless the quality of such data is controlled, more training data can cause translation performance to decrease drastically, by up to 38% relative BLEU in our experiments. 1 Introduction Statistical machine translation (SMT) systems are trained on sentence-aligned parallel corpora con- sisting of translated texts. In the simplest case the translation direction is constant so that one part of the parallel corpus is the translation of the other. In more complex cases, either some texts may have been translated from language A to lan- guage B and others the other way round, or more than two languages are involved and both parts were translated from one another or several other languages. This is the case of corpora involving European languages, such as the Europarl corpus c⃝ 2009 European Association for Machine Translation. (Koehn, 2005)1 or the Acquis Communautaire cor- pus (Steinberger et al., 2006)2, which comprise texts coming from institutions of the European Union. They are amongst the largest and most widely used corpora in SMT. Typically, given a corpus in language A, its ver- sion in language B and an SMT system translating from A to B, SMT training assumes A to be the source language (SL) and B to be the target lan- guage (TL) irrespective of the original translation direction or languages involved. In other words, it is assumed that the original SL does not matter when training an SMT system which aims to trans- late from language A to language B. Following a brief overview of related work (sec- tion 2), we investigate the impact of the original SL with regard to French–English translation. Our experimental objective is to compare training con- figurations which differ in terms of the original SL by measuring French-to-English and English- to-French translation quality of a state-of-the-art phrase-based SMT (PB-SMT) system. We train four different configurations of the same PB-SMT system based on French–English parallel corpora which differ in terms of the original SL (sections 3 and 4) and carry out translation experiments from French into English and from English into French (section 5). We evaluate each output using stan- dard evaluation metrics, compare the results and present our findings (section …

Content truncated to 500 words