The JAIST Machine Translation Systems for WMT 17

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 405–409 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics The JAIST Machine Translation Systems for WMT 17 Hai-Long Trieu and Trung-Tin Pham and Le-Minh Nguyen School of Information Science Japan Advanced Institute of Science and Technology {trieulh, tinpt, nguyenml}@jaist.ac.jp Abstract We describe the JAIST phrase-based ma- chine translation systems that participated in the news translation shared task of the WMT17. In this work, we participated in the Turkish-English translation, in which only a small amount of bilingual training data is available, so that it is an exam- ple of the low-resource setting in machine translation. In order to solve the prob- lem, we focus on two strategies: build- ing a bilingual corpus from comparable data and exploiting existing parallel data based on phrase pivot translation. In or- der to utilize the strategies to enhance ma- chine translation on the low-resource set- ting most effectively, we introduce a sys- tem combining the extracted corpus, the pivot translation, and the direct training data. Experimental results showed that our combined systems significantly improved the baseline models, which were trained on the small bilingual data. 1 Introduction We participated in the WMT 17 news translation shared task for the Turkish-English language pair. The amount of bilingual training data for this lan- guage pair is small, which means that this ma- chine translation task poses the problem of a low- resource setting. The problem causes a bottleneck for current data-driven machine translation meth- ods including phrase-based and neural-based ma- chine translation because there are few large bilin- gual corpora for most language pairs in the world (Irvine, 2013; Wang et al., 2016). In our systems, we focus on two strategies to enhance machine translation for the low-resource setting: building a bilingual corpus from compa- rable data, and exploiting existing parallel cor- pora based on the phrase pivot translation (Wu and Wang, 2007; Cohn and Lapata, 2007; Utiyama and Isahara, 2007). First, we built a bilingual cor- pus for Turkish-English based on parallel titles of Wikipedia articles. The parallel titles were ex- tracted from Wikipedia articles’ titles and inter- language link records. Bilingual articles were col- lected based on the title pairs. Then, bilingual sentences were extracted from the article pairs us- ing the Microsoft sentence aligner (Moore, 2002). Second, we exploited the phrase pivot translation method using six pivot languages to bridge the translation between Turkish and English. Finally, the two resources of the extracted corpus and the pivot translation were utilized with the direct bilin- gual training data in a combined system. Our combined systems achieved a significant improve- ment compared with the baseline model, which was trained on the direct bilingual data. The code and datasets used in our systems can be found at the repository.1 2 Methods We describe approaches used in our systems. The Turkish-English bilingual data in this shared task embodies only 207k parallel sentences, which is an instance of machine translation task in …

Content truncated to 500 words