Improving Pivot-Based Statistical Machine Translation by Pivoting the Co-occurrence Count of Phrase Pairs

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1665–1675, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Improving Pivot-Based Statistical Machine Translation by Pivoting the Co-occurrence Count of Phrase Pairs Xiaoning Zhu1*, Zhongjun He2, Hua Wu2, Conghui Zhu1, Haifeng Wang2, and Tiejun Zhao1 Harbin Institute of Technology, Harbin, China1 {xnzhu,chzhu,tjzhao}@mtlab.hit.edu.cn Baidu Inc., Beijing, China2 {hezhongjun,wu_hua,wanghaifeng}@baidu.com * This work was done when the first author was visiting Baidu. Abstract To overcome the scarceness of bilingual corpora for some language pairs in ma- chine translation, pivot-based SMT uses pivot language as a "bridge" to generate source-target translation from source- pivot and pivot-target translation. One of the key issues is to estimate the probabili- ties for the generated phrase pairs. In this paper, we present a novel approach to calculate the translation probability by pivoting the co-occurrence count of source-pivot and pivot-target phrase pairs. Experimental results on Europarl data and web data show that our method leads to significant improvements over the baseline systems. 1 Introduction Statistical Machine Translation (SMT) relies on large bilingual parallel data to produce high qual- ity translation results. Unfortunately, for some language pairs, large bilingual corpora are not readily available. To alleviate the parallel data scarceness, a conventional solution is to intro- duce a “bridge” language (named pivot language) to connect the source and target language (de Gispert and Marino, 2006; Utiyama and Isahara, 2007; Wu and Wang, 2007; Bertoldi et al., 2008; Paul et al., 2011; El Kholy et al., 2013; Zahabi et al., 2013), where there are large amounts of source-pivot and pivot-target parallel corpora. Among various pivot-based approaches, the triangulation method (Cohn and Lapata, 2007; Wu and Wang, 2007) is a representative work in pivot-based machine translation. The approach proposes to build a source-target phrase table by merging the source-pivot and pivot-target phrase table. One of the key issues in this method is to estimate the translation probabilities for the gen- erated source-target phrase pairs. Conventionally, the probabilities are estimated by multiplying the posterior probabilities of source-pivot and pivot- target phrase pairs. However, it has been shown that the generated probabilities are not accurate enough (Cui et al., 2013). One possible reason may lie in the non-uniformity of the probability space. Through Figure 1. (a), we can see that the probability distributions of source-pivot and piv- ot-target language are calculated separately, and the source-target probability distributions are induced from the source-pivot and pivot-target probability distributions. Because of the absence of the pivot language (e.g., p2 is in source-pivot probability space but not in pivot-target one), the induced source-target probability distribution is not complete, which will result in inaccurate probabilities. To solve this problem, we propose a novel ap- proach that utilizes the co-occurrence count of source-target phrase pairs to estimate phrase translation probabilities more precisely. Different from the triangulation method, which merges the source-pivot and pivot-target phrase pairs after training the translation model, we propose to merge the source-pivot and pivot-target phrase pairs immediately after the phrase …

Content truncated to 500 words