Phrase Model Training for Statistical Machine Translation with Word Lattices of Preprocessing Alternatives

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 450–459, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Phrase Model Training for Statistical Machine Translation with Word Lattices of Preprocessing Alternatives Joern Wuebker and Hermann Ney Human Language Technology and Pattern Recognition Group RWTH Aachen University Aachen, Germany {wuebker,ney}@cs.rwth-aachen.de Abstract In statistical machine translation, word lattices are used to represent the ambiguities in the preprocessing of the source sentence, such as word segmentation for Chinese or morpholog- ical analysis for German. Several approaches have been proposed to define the probability of different paths through the lattice with ex- ternal tools like word segmenters, or by apply- ing indicator features. We introduce a novel lattice design, which explicitly distinguishes between different preprocessing alternatives for the source sentence. It allows us to make use of specific features for each preprocess- ing type and to lexicalize the choice of lattice path directly in the phrase translation model. We argue that forced alignment training can be used to learn lattice path and phrase trans- lation model simultaneously. On the news- commentary portion of the German→English WMT 2011 task we can show moderate im- provements of up to 0.6% BLEU over a state- of-the-art baseline system. 1 Introduction The application of statistical machine translation (SMT) to word lattice input was first introduced for the translation of speech recognition output. Rather than translating the single-best transcription, the speech recognition system encodes all possible tran- scriptions and their probabilities within a word lat- tice, which is then used as input for the machine translation system (Ney, 1999; Matusov et al., 2005; Bertoldi et al., 2007). Since then, several groups have adapted this ap- proach to model ambiguities in representing the source language with lattices and were able to re- port improvements over their respective baselines. The probabilities for different paths through the lat- tice are usually modeled by assigning probabilities to arcs as a byproduct of the lattice generation or by defining binary indicator features. Applying the first method only makes sense if the lattice construc- tion is based on a single, comprehensive probabilis- tic method, like a Chinese word segmentation model as is used by Xu et al. (2005). In applications like the one described by Dyer et al. (2008), where sev- eral different segmenters for Chinese are combined to create the lattice, this is not possible. Also, our intuition suggests that simply defining indicator fea- tures for each of the segmenters may not be ideal, if we assume that there is not a single best segmenter, but rather that for different data instances a different one works best. In this paper, we propose to model the lattice path implicitly within the phrase translation model. We introduce a novel lattice design, which explic- itly distinguishes between different ways of prepro- cessing the source sentence. It enables us to define specific binary features for each preprocessing type and to learn lexicalized lattice path probabilities and the phrase translation model simultaneously with a forced alignment …

Content truncated to 500 words