Computing Consensus Translation for Multiple Machine Translation Systems Using Enhanced Hypothesis Alignment

Computing Consensus Translation from Multiple Machine Translation Systems Using Enhanced Hypotheses Alignment Evgeny Matusov, Nicola Ueffing, Hermann Ney Lehrstuhl f¨ur Informatik VI - Computer Science Department RWTH Aachen University, Aachen, Germany. {matusov,ueffing,ney}@informatik.rwth-aachen.de Abstract This paper describes a novel method for computing a consensus translation from the outputs of multiple machine trans- lation (MT) systems. The outputs are combined and a possibly new transla- tion hypothesis can be generated. Simi- larly to the well-established ROVER ap- proach of (Fiscus, 1997) for combining speech recognition hypotheses, the con- sensus translation is computed by voting on a confusion network. To create the con- fusion network, we produce pairwise word alignments of the original machine trans- lation hypotheses with an enhanced sta- tistical alignment algorithm that explicitly models word reordering. The context of a whole document of translations rather than a single sentence is taken into account to produce the alignment. The proposed alignment and voting ap- proach was evaluated on several machine translation tasks, including a large vocab- ulary task. The method was also tested in the framework of multi-source and speech translation. On all tasks and conditions, we achieved significant improvements in translation quality, increasing e. g. the BLEU score by as much as 15% relative. 1 Introduction In this work we describe a novel technique for computing a consensus translation from the out- puts of multiple machine translation systems. Combining outputs from different systems was shown to be quite successful in automatic speech recognition (ASR). Voting schemes like the ROVER approach of (Fiscus, 1997) use edit distance alignment and time information to cre- ate confusion networks from the output of several ASR systems. Some research on multi-engine machine trans- lation has also been performed in recent years. The most straightforward approaches simply se- lect, for each sentence, one of the provided hy- potheses. The selection is made based on the scores of translation, language, and other mod- els (Nomoto, 2004; Paul et al., 2005). Other approaches combine lattices or N-best lists from several different MT systems (Frederking and Nirenburg, 1994). To be successful, such ap- proaches require compatible lattices and compa- rable scores of the (word) hypotheses in the lat- tices. However, the scores of most statistical ma- chine translation (SMT) systems are not normal- ized and therefore not directly comparable. For some other MT systems (e.g. knowledge-based systems), the lattices and/or scores of hypotheses may not be even available. (Bangalore et al., 2001) used the edit distance alignment extended to multiple sequences to con- struct a confusion network from several transla- tion hypotheses. This algorithm produces mono- tone alignments only (i. e. allows insertion, dele- tion, and substitution of words); it is not able to align translation hypotheses with significantly dif- ferent word order. (Jayaraman and Lavie, 2005) try to overcome this problem. They introduce a method that allows non-monotone alignments of words in different translation hypotheses for the same sentence. However, this approach uses many heuristics and is based on the alignment that is per- formed to calculate a …

Content truncated to 500 words