Revisiting Pivot Language Approach for Machine Translation

Proceedings of the 47th Annual Meeting of the ACL and the 4th IJCNLP of the AFNLP, pages 154–162, Suntec, Singapore, 2-7 August 2009. c⃝2009 ACL and AFNLP Revisiting Pivot Language Approach for Machine Translation Hua Wu and Haifeng Wang Toshiba (China) Research and Development Center 5/F., Tower W2, Oriental Plaza, Beijing, 100738, China {wuhua, wanghaifeng}@rdc.toshiba.com.cn Abstract This paper revisits the pivot language ap- proach for machine translation. First, we investigate three different methods for pivot translation. Then we employ a hybrid method combining RBMT and SMT systems to fill up the data gap for pivot translation, where the source- pivot and pivot-target corpora are inde- pendent. Experimental results on spo- ken language translation show that this hybrid method significantly improves the translation quality, which outperforms the method using a source-target corpus of the same size. In addition, we pro- pose a system combination approach to select better translations from those pro- duced by various pivot translation meth- ods. This method regards system com- bination as a translation evaluation prob- lem and formalizes it with a regression learning model. Experimental results in- dicate that our method achieves consistent and significant improvement over individ- ual translation outputs. 1 Introduction Current statistical machine translation (SMT) sys- tems rely on large parallel and monolingual train- ing corpora to produce translations of relatively higher quality. Unfortunately, large quantities of parallel data are not readily available for some lan- guages pairs, therefore limiting the potential use of current SMT systems. In particular, for speech translation, the translation task often focuses on a specific domain such as the travel domain. It is es- pecially difficult to obtain such a domain-specific corpus for some language pairs such as Chinese to Spanish translation. To circumvent the data bottleneck, some re- searchers have investigated to use a pivot language approach (Cohn and Lapata, 2007; Utiyama and Isahara, 2007; Wu and Wang 2007; Bertoldi et al., 2008). This approach introduces a third language, named the pivot language, for which there exist large source-pivot and pivot-target bilingual cor- pora. A pivot task was also designed for spoken language translation in the evaluation campaign of IWSLT 2008 (Paul, 2008), where English is used as a pivot language for Chinese to Spanish trans- lation. Three different pivot strategies have been in- vestigated in the literature. The first is based on phrase table multiplication (Cohn and Lap- ata 2007; Wu and Wang, 2007). It multiples corresponding translation probabilities and lexical weights in source-pivot and pivot-target transla- tion models to induce a new source-target phrase table. We name it the triangulation method. The second is the sentence translation strategy, which first translates the source sentence to the pivot sen- tence, and then to the target sentence (Utiyama and Isahara, 2007; Khalilov et al., 2008). We name it the transfer method. The third is to use existing models to build a synthetic source-target corpus, from which a source-target model can be trained (Bertoldi et al., 2008). For example, we can ob- tain a source-pivot corpus by translating …

Content truncated to 500 words