Distortion Models for Statistical Machine Translation

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 529–536, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Distortion Models For Statistical Machine Translation Yaser Al-Onaizan and Kishore Papineni IBM T.J. Watson Research Center 1101 Kitchawan Road Yorktown Heights, NY 10598, USA {onaizan, papineni}@us.ibm.com Abstract In this paper, we argue that n-gram lan- guage models are not sufficient to address word reordering required for Machine Trans- lation. We propose a new distortion model that can be used with existing phrase-based SMT decoders to address those n-gram lan- guage model limitations. We present empirical results in Arabic to English Machine Transla- tion that show statistically significant improve- ments when our proposed model is used. We also propose a novel metric to measure word order similarity (or difference) between any pair of languages based on word alignments. 1 Introduction A language model is a statistical model that gives a probability distribution over possible sequences of words. It computes the probability of producing a given word w1 given all the words that precede it in the sen- tence. An n-gram language model is an n-th order Markov model where the probability of generating a given word depends only on the last n − 1 words im- mediately preceding it and is given by the following equation: P(wk 1) = P(w1)P(w2|w1) · · · P(wn|wn−1 1 ) (1) where k >= n. N-gram language models have been successfully used in Automatic Speech Recognition (ASR) as was first proposed by (Bahl et al., 1983). They play an im- portant role in selecting among several candidate word realization of a given acoustic signal. N-gram lan- guage models have also been used in Statistical Ma- chine Translation (SMT) as proposed by (Brown et al., 1990; Brown et al., 1993). The run-time search pro- cedure used to find the most likely translation (or tran- scription in the case of Speech Recognition) is typically referred to as decoding. There is a fundamental difference between decoding for machine translation and decoding for speech recog- nition. When decoding a speech signal, words are gen- erated in the same order in which their corresponding acoustic signal is consumed. However, that is not nec- essarily the case in MT due to the fact that different languages have different word order requirements. For example, in Spanish and Arabic adjectives are mainly noun post-modifiers, whereas in English adjectives are noun pre-modifiers. Therefore, when translating be- tween Spanish and English, words must usually be re- ordered. Existing statistical machine translation decoders have mostly relied on language models to select the proper word order among many possible choices when translating between two languages. In this paper, we argue that a language model is not sufficient to ade- quately address this issue, especially when translating between languages that have very different word orders as suggested by our experimental results in Section 5. We propose a new distortion model that can be used as an additional component in SMT decoders. This …

Content truncated to 500 words