Improving Chinese-to-Japanese Patent Translation Using English as Pivot Language

Copyright 2012 by Xianhua Li, Yao Meng, and Yao Meng 26th Pacific Asia Conference on Language,Information and Computation pages 117–126 Improving Chinese-to-Japanese Patent Translation Using English as Pivot Language Xianhua Li Yao Meng Hao Yu Fujitsu R&D Centre CO., LTD, Beijing, China {lixianhua, mengyao, yu}@cn.fujitsu.com Abstract This paper implements and compares three different strategies to use English as pivot language for Chinese-Japanese patent translation: corpus enrichment, sentence pivot translation and phrase pivot translation. Our results show that both corpus enrichment and phrase pivot translation strategy outperform the baseline system, while the sentence pivot translation strategy failed to improve the system. We apply the strategies on large data set and figure out approaches to improve efficiency. Finally, we perform Minimum Bayes Risk system combination on the different results of direct translation system and pivot translation systems, which significantly outperforms the direct translation system by 4.25 BLEU scores. 1 Introduction Statistical machine translation (SMT) has made rapid progress in recent years with the support of large quantities of parallel corpora. It’s quite common that we use millions of bilingual parallel sentences to train a statistical machine translation system. Unfortunately, large parallel corpora are not always available for some language pairs, or for some specific domains. For example, there are few available bilingual corpora for Chinese-to- Japanese patent translation. Many research labs and companies face data bottleneck when they do research on scare-resourced language pairs or domains. Much work has been done to overcome the data bottleneck problem. For example, Lu et al. (2009) exploited the existence of bilingual patent corpora and constructed a Chinese-English patent parallel corpus. Resnik and Smith (2003) took the web as a parallel corpus and mined parallel data from it. Munteanu and Marcu (2005) trained a maximum entropy classifier to extract parallel corpus from large non-parallel newspaper corpora. Our work differs in that we make use of the currently available bilingual corpora, without exploiting extra bilingual data to improve machine translation quality. In other words, we employ pivot translation strategies to improve the performance of SMT systems.  How to apply pivot translation strategies to help scare-resourced language translation?  How to take advantages of different pivot translation strategies to further improve machine translation quality? In this paper, we introduce and implement three pivot translation strategies for SMT. The first is corpus enrichment strategy. It translates the pivot side of source-pivot corpus and pivot-target corpus into target and source language respectively to construct source-target language pairs. With these sentence pairs, it builds up a new SMT system so as to outperform the basic system. As the corpora we employ are quite large, we select sentence pairs according to their sentence value and do experiments on different size of parallel corpus. The second is sentence pivot translation strategy. 117 Copyright 2012 by Xianhua Li, Yao Meng and Hao Yu 26 th Pacific Asia Conference on Language, Information and Computation pages 117-126 It builds two SMT systems on source-pivot and pivot-target corpus respectively. When translating a source sentence into target …

Content truncated to 500 words