NUT-NTT Statistical Machine Translation System for IWSLT 2005

NUT-NTT Statistical Machine Translation System for IWSLT 2005 Kazuteru Ohashi and Kazuhide Yamamoto Nagaoka University of Technology 1603-1, Kamitomioka, Nagaoka City Niigata, 940-2188 Japan � ohashi, ykaz � @nlp.nagaokaut.ac.jp Kuniko Saito and Masaaki Nagata NTT Cyber Space Laboratories 1-1 Hikarinooka, Yokoshuka-shi Kanagawa, 239-0847 Japan � saito.kuniko, masaaki.nagata � @labs.ntt.co.jp Abstract In this paper, we present a novel distortion model for phrase-based statistical machine translation. Unlike the pre- vious phrase distortion models whose role is to simply penal- ize nonmonotonic alignments[1, 2], the new model assigns the probability of relative position between two source lan- guage phrases aligned to the two adjacent target language phrases. The phrase translation probabilities and phrase dis- tortion probabilities are calculated from the N-best phrase alignment of the training bilingual sentences. To obtain N- best phrase alignment, we devised a novel phrase alignment algorithm based on word translation probabilities and N-best search. Experiments show that the phrase distortion model and phrase translation model improve the BLEU and NIST scores over the baseline method. 1. Introduction In recent years, phrase-based translation models have be- come the mainstream of statistical machine translation, be- cause they can represent context-based word selection and local word reordering better than word-based translation models. Previous phrased-based translation models[1, 2], however, are not effective for global phrase reordering, be- cause their distortion model is too simplistic. As it was de- signed simply to penalize nonmonotonic phrase alignment, it is difficult to handle translations that require complex word reordering, such as between Japanese and English. In this paper, we present a novel distortion model for phrase-based statistical machine translation. It models the probability of relative position between two source language phrases aligned to the two adjacent target language phrases. To obtain the distortion model, we first make a phrase align- ment of each sentence pair in the training corpus. We then calculate the phrase distortion probability from the relative frequency of respective events in the phrase aligned train- ing corpus. In order to cope with the sparse data prob- lem, word reordering is classified into four states: mono- tone, monotone-gap, reverse, and reverse-gap. Phrases are also classified based on the part of speech of the first and last word. We need phrase translation probabilities to get phrase alignment from the training corpus, but we need phrase align- ment to get phrase translation probabilities. To solve this chicken and egg problem, we devised a novel phrase align- ment algorithm using word translation probabilities and for- ward beam search. Phrase distortion probabilities mentioned above are calculated from the result of this phrase alignment. The phrase alignment algorithm can easily be extended to obtain N-best phrase alignment using backward A* search, such as [3]. We found that phrase translation probabilities calculated from the result of this N-best phrase alignment improve the translation accuracy significantly. In the following sections, we first explain our transla- tion model including the phrase distortion model and phrase alignment algorithm. We then report the experiments’ results and show the effectiveness of our phrase distortion …

Content truncated to 500 words