Improving Pivot-Based Statistical Machine Translation Using Random Walk

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 524–534, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Improving Pivot-Based Statistical Machine Translation Using Random Walk Xiaoning Zhu1* Conghui Zhu1, and Tiejun Zhao1 , Zhongjun He2, Hua Wu2, Haifeng Wang2, Harbin Institute of Technology, Harbin, China1 Baidu Inc., Beijing, China2 {xnzhu, chzhu, tjzhao}@mtlab.hit.edu.cn {hezhongjun,wu_hua,wanghaifeng}@baidu.com * This work was done when the first author was visiting Baidu. Abstract This paper proposes a novel approach that uti- lizes a machine learning method to improve pivot-based statistical machine translation (SMT). For language pairs with few bilingual data, a possible solution in pivot-based SMT using another language as a "bridge" to gen- erate source-target translation. However, one of the weaknesses is that some useful source- target translations cannot be generated if the corresponding source phrase and target phrase connect to different pivot phrases. To allevi- ate the problem, we utilize Markov random walks to connect possible translation phrases between source and target language. Experi- mental results on European Parliament data, spoken language data and web data show that our method leads to significant improvements on all the tasks over the baseline system. 1 Introduction Statistical machine translation (SMT) uses bilin- gual corpora to build translation models. The amount and the quality of the bilingual data strongly affect the performance of SMT systems. For resource-rich language pairs, such as Chinese- English, it is easy to collect large amounts of bi- lingual corpus. However, for resource-poor lan- guage pairs, such as Chinese-Spanish, it is difficult to build a high-performance SMT system with the small scale bilingual data available. The pivot language approach, which performs translation through a third language, provides a possible solution to the problem. The triangulation method (Wu and Wang, 2007; Cohn and Lapata, 2007) is a representative work for pivot-based ma- chine translation. With a triangulation pivot ap- proach, a source-target phrase table can be obtained by combining the source-pivot phrase table and the pivot-target phrase table. However, one of the weaknesses is that some corresponding source and target phrase pairs cannot be generated, because they are connected to different pivot phrases (Cui et al., 2013). As illustrated in Figure 1, since there is no direct translation between “很 可口 henkekou” and “really delicious”, the trian- gulation method is unable to establish a relation between “很可口 henkekou” and the two Spanish phrases. To solve this problem, we apply a Markov ran- dom walk method to pivot-based SMT system. Random walk has been widely used. For example, Brin and Page (1998) used random walk to dis- cover potential relations between queries and doc- uments for link analysis in information retrieval. Analogous to link analysis, the aim of pivot-based translation is to discover potential translations be- tween source and target language via the pivot language. 524 The goal of this paper is to extend the previous triangulation approach by exploring implicit trans- lation relations using random walk method. We evaluated our approach in several translation tasks, including translations …

Content truncated to 500 words