Multi-Engine Machine Translation by Recursive Sentence Decomposition

Multi-Engine Machine Translation by Recursive Sentence Decomposition Bart Mellebeek School of Computing Dublin City University Dublin 9, Ireland mellebeek@computing.dcu.ie Karolina Owczarzak School of Computing Dublin City University Dublin 9, Ireland owczarzak@computing.dcu.ie Josef Van Genabith School of Computing Dublin City University Dublin 9, Ireland josef@computing.dcu.ie Andy Way School of Computing Dublin City University Dublin 9, Ireland away@computing.dcu.ie Abstract In this paper, we present a novel approach to combine the outputs of multiple MT en- gines into a consensus translation. In con- trast to previous Multi-Engine Machine Translation (MEMT) techniques, we do not rely on word alignments of output hy- potheses, but prepare the input sentence for multi-engine processing. We do this by using a recursive decomposition algo- rithm that produces simple chunks as in- put to the MT engines. A consensus trans- lation is produced by combining the best chunk translations, selected through ma- jority voting, a trigram language model score and a confidence score assigned to each MT engine. We report statis- tically significant relative improvements of up to 9% BLEU score in experiments (English→Spanish) carried out on an 800- sentence test set extracted from the Penn- II Treebank. 1 Introduction In this paper, we present a novel approach to com- bine the outputs of multiple MT engines into a con- sensus translation. Multi-Engine Machine Translation (MEMT) is a term coined by (Frederking and Nirenburg, 1994), who were the first to apply the idea of a multi- engine approach in Natural Language Processing to Machine Translation (MT). Researchers in other ar- eas of language technology such as Speech Recog- nition (Fiscus, 1997), Text Categorization (Larkey and Croft, 1996) and POS Tagging (Roth and Ze- lenko, 1998) have also experimented with multi- system approaches. Since then, several researchers in the MT community have come up with different techniques to calculate consensus translations from multiple MT engines (cf. section 2). All these previ- ously proposed techniques share one important char- acteristic: they translate the entire input sentence as is and operate on the resulting target language sen- tences to calculate a consensus output. Their main difference lies in the method they use to compute word alignments between the multiple output sen- tences. In contrast to previous MEMT approaches, the technique we present does not rely on word align- ments of target language sentences, but is based on a recursive chunking algorithm that produces simple constituents as input to the MT engines. The outputs of these syntactically meaningful chunks are com- pared to each other and the highest ranked transla- tions are used to compose the output sentence. Our approach, therefore, prepares the input sentence for multi-engine processing on the input side. It draws its strength from the simple fact that short input strings result in better translations than longer ones. The paper is organised as follows. In section 2, we 110 Proceedings of the 7th Conference of the Association for Machine Translation in the Americas, pages 110-118, Cambridge, August 2006. ©2006 The Association for Machine Translation in the Americas give a short …

Content truncated to 500 words