Improving Sampling-based Alignment by Investigating the Distribution of N-grams in Phrase Translation Tables

Improving Sampling-based Alignment by Investigating the Distribution of N-grams in Phrase Translation Tables ⋆ Juan Luo a, Adrien Lardilleux b, and Yves Lepage a aGraduate School of Information, Production and Systems, Waseda University, 2-7 Hibikino, Wakamatsu-ku, Kitakyushu, Fukuoka 808-0135, Japan {juanluoonly@suou, yves.lepage@aoni}.waseda.jp bTLP Group, LIMSI-CNRS, BP 133, Orsay Cedex 91403, France adrien.lardilleux@limsi.fr Abstract. This paper describes an approach to improve the performance of sampling-based multilingual alignment on translation tasks by investigating the distribution of n-grams in the translation tables. This approach consists in enforcing the alignment of n-grams. The quality of phrase translation tables output by this approach and that of MGIZA++ is compared in statistical machine translation tasks. Significant improvements for this approach are reported. In addition, merging translation tables is shown to outperform state-of-the-art techniques. Keywords: alignment, phrase translation table, statistical machine translation. 1 Introduction Phrase translation tables play an important role in the process of building machine translation systems. The quality of translation table, which identifies the relations between words or phrases in the source language and those in the target language, is crucial for the quality of the output of most machine translation systems. Currently, the most widely used state-of-the-art tool to generate phrase translation tables is GIZA++ (Och and Ney, 2003), which trains the ubiquitous IBM models (Brown et al., 1993) and the HMM introduced by (Vogel et al., 1996), in combination with the Moses toolkit (Koehn et al., 2007). MGIZA++, a multi-threaded word aligner based on GIZA++, is proposed by (Gao and Vogel, 2008). In this paper, we investigate a different approach to the production of phrase translation tables: the sampling-based approach (Lardilleux and Lepage, 2009b). This approach is implemented in a free open-source tool called Anymalign.1 Being in line with the associative alignment trend illustrated by (Gale and Church, 1991; Melamed, 2000; Moore, 2005), it is much simpler than the models implemented in MGIZA++, which are in line with the estimating trend illustrated by (Brown et al., 1991; Och and Ney, 2003; Liang et al., 2006). In addition, it is capable of aligning multiple languages simultaneously; but we will not use this feature here as we will restrain ourselves to bilingual experiments in this paper. In sampling-based alignment, only those sequences of words sharing the exact same distribu- tion (i.e., they appear exactly in the same sentences of the corpus) are considered for alignment. ⋆ Part of the research presented in this paper has been done under a Japanese grant-in-aid (Kakenhi C, A11515600: Improvement of alignments and release of multilingual syntactic patterns for statistical and example-based machine translation). Copyright 2011 by Juan Luo, Adrien Lardilleux, and Yves Lepage 1 http://users.info.unicaen.fr/˜alardill/anymalign/ 25th Pacific Asia Conference on Language, Information and Computation, pages 150–159 150 The key idea is to make more words share the same distribution by artificially reducing their fre- quency in multiple random subcorpora obtained by sampling. Indeed, the smaller a subcorpus, the less frequent its words, and the more likely they are to share the same distribution; hence the higher the proportion …

Content truncated to 500 words