N-Gram Posterior Probabilities for Statistical Machine Translation

Proceedings of the Workshop on Statistical Machine Translation, pages 72–77, New York City, June 2006. c⃝2006 Association for Computational Linguistics N-Gram Posterior Probabilities for Statistical Machine Translation Richard Zens and Hermann Ney Human Language Technology and Pattern Recognition Lehrstuhl f¨ur Informatik 6 – Computer Science Department RWTH Aachen University, D-52056 Aachen, Germany {zens,ney}@cs.rwth-aachen.de Abstract Word posterior probabilities are a com- mon approach for confidence estimation in automatic speech recognition and ma- chine translation. We will generalize this idea and introduce n-gram posterior prob- abilities and show how these can be used to improve translation quality. Addition- ally, we will introduce a sentence length model based on posterior probabilities. We will show significant improvements on the Chinese-English NIST task. The abso- lute improvements of the BLEU score is between 1.1% and 1.6%. 1 Introduction The use of word posterior probabilities is a com- mon approach for confidence estimation in auto- matic speech recognition, e.g. see (Wessel, 2002). This idea has been adopted to estimate confidences for machine translation, e.g. see (Blatz et al., 2003; Ueffing et al., 2003; Blatz et al., 2004). These confi- dence measures were used in the computer assisted translation (CAT) framework, e.g. (Gandrabur and Foster, 2003). The (simplified) idea is that the con- fidence measure is used to decide if the machine- generated prediction should be suggested to the hu- man translator or not. There is only few work on how to improve machine translation performance using confidence measures. The only work, we are aware of, is (Blatz et al., 2003). The outcome was that the con- fidence measures did not result in improvements of the translation quality measured with the BLEU and NIST scores. Here, we focus on how the ideas and methods commonly used for confidence estimation can be adapted and/or extended to improve transla- tion quality. So far, always word-level posterior probabilities were used. Here, we will generalize this idea to n- grams. In addition to the n-gram posterior probabili- ties, we introduce a sentence-length model based on posterior probabilities. The common phrase- based translation systems, such as (Och et al., 1999; Koehn, 2004), do not use an explicit sentence length model. Only the simple word penalty goes into that direction. It can be adjusted to prefer longer or shorter translations. Here, we will explicitly model the sentence length. The novel contributions of this work are to in- troduce n-gram posterior probabilities and sentence length posterior probabilities. Using these methods, we achieve significant improvements of translation quality. The remaining part of this paper is structured as follows: first, we will briefly describe the baseline system, which is a state-of-the-art phrase-based sta- tistical machine translation system. Then, in Sec- tion 3, we will introduce the n-gram posterior prob- abilities. In Section 4, we will define the sentence length model. Afterwards, in Section 5, we will describe how these novel models can be used for rescoring/reranking. The experimental results will be presented in Section 6. Future applications will be described in Section 7. Finally, we will …

Content truncated to 500 words