Improvements in Phrase-Based Statistical Machine Translation

Improvements in Phrase-Based Statistical Machine Translation Richard Zens and Hermann Ney Chair of Computer Science VI RWTH Aachen University {zens,ney}@cs.rwth-aachen.de Abstract In statistical machine translation, the currently best performing systems are based in some way on phrases or word groups. We describe the baseline phrase-based translation system and various refinements. We describe a highly ef- ficient monotone search algorithm with a com- plexity linear in the input sentence length. We present translation results for three tasks: Verb- mobil, Xerox and the Canadian Hansards. For the Xerox task, it takes less than 7 seconds to translate the whole test set consisting of more than 10K words. The translation results for the Xerox and Canadian Hansards task are very promising. The system even outperforms the alignment template system. 1 Introduction In statistical machine translation, we are given a source language (‘French’) sentence f J 1 = f1 . . . fj . . . fJ, which is to be translated into a target language (‘English’) sentence eI 1 = e1 . . . ei . . . eI. Among all possible target language sentences, we will choose the sentence with the highest probability: ˆeI 1 = argmax eI 1 � Pr(eI 1|f J 1 ) � (1) = argmax eI 1 � Pr(eI 1) · Pr(f J 1 |eI 1) � (2) The decomposition into two knowledge sources in Equa- tion 2 is known as the source-channel approach to statisti- cal machine translation (Brown et al., 1990). It allows an independent modeling of target language model Pr(eI 1) and translation model Pr(f J 1 |eI 1)1. The target language 1The notational convention will be as follows: we use the symbol Pr(·) to denote general probability distributions with (nearly) no specific assumptions. In contrast, for model-based probability distributions, we use the generic symbol p(·). model describes the well-formedness of the target lan- guage sentence. The translation model links the source language sentence to the target language sentence. It can be further decomposed into alignment and lexicon model. The argmax operation denotes the search problem, i.e. the generation of the output sentence in the target lan- guage. We have to maximize over all possible target lan- guage sentences. An alternative to the classical source-channel ap- proach is the direct modeling of the posterior probabil- ity Pr(eI 1|f J 1 ). Using a log-linear model (Och and Ney, 2002), we obtain: Pr(eI 1|f J 1 ) = exp � M � m=1 λmhm(eI 1, f J 1 ) � · Z(f J 1 ) Here, Z(f J 1 ) denotes the appropriate normalization con- stant. As a decision rule, we obtain: ˆeI 1 = argmax eI 1 � M � m=1 λmhm(eI 1, f J 1 ) � This approach is a generalization of the source-channel approach. It has the advantage that additional models or feature functions can be easily integrated into the overall system. The model scaling factors λM 1 are trained accord- ing to the maximum entropy principle, e.g. using the GIS algorithm. …

Content truncated to 500 words