Leveraging Small Multilingual Corpora for SMT Using Many Pivot Languages

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1192–1202, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Leveraging Small Multilingual Corpora for SMT Using Many Pivot Languages Raj Dabre Graduate School of Informatics Kyoto University Kyoto 606-8501 prajdabre@gmail.com Fabien Cromieres Japan Science and Technology Agency Kawaguchi-shi Saitama 332-0012 fabien@pa.jst.jp Sadao Kurohashi Graduate School of Informatics Kyoto University Kyoto 606-8501 kuro@i.kyoto-u.ac.jp Pushpak Bhattacharyya CFILT IIT Bombay, Powai India 400076 pushpakbh@gmail.com Abstract We present our work on leveraging multilin- gual parallel corpora of small sizes for Sta- tistical Machine Translation between Japanese and Hindi using multiple pivot languages. In our setting, the source and target part of the corpus remains the same, but we show that using several different pivot to extract phrase pairs from these source and target parts lead to large BLEU improvements. We focus on a variety of ways to exploit phrase tables gener- ated using multiple pivots to support a direct source-target phrase table. Our main method uses the Multiple Decoding Paths (MDP) fea- ture of Moses, which we empirically verify as the best compared to the other methods we used. We compare and contrast our various re- sults to show that one can overcome the limita- tions of small corpora by using as many pivot languages as possible in a multilingual setting. Most importantly, we show that such pivot- ing aids in learning of additional phrase pairs which are not learned when the direct source- target corpus is small. We obtained improve- ments of up to 3 BLEU points using multiple pivots for Japanese to Hindi translation com- pared to when only one pivot is used. To the best of our knowledge, this work is also the first of its kind to attempt the simultaneous uti- lization of 7 pivot languages at decoding time. 1 Introduction With the increasing size of parallel corpora it has become possible to achieve very high quality trans- lation. However, not all language pairs are blessed with the availability of large parallel corpora in the sizes of millions of lines. With the exception of the major European languages and a few Asian lan- guages like Chinese and Japanese, other languages have parallel corpora in the sizes of a few thousands of lines. Since translation quality is related to the size of the parallel corpus, it is impossible to achieve the same level of translation quality as that in the case of resource rich languages. To remedy this sce- nario, an intermediate resource rich language can be exploited. Although, finding a direct parallel cor- pus between source and target languages might be difficult, there are higher odds of finding a pair of parallel corpora: one between the source language and an intermediate resource rich language (hence- forth called pivot1) and one between that pivot and the target language. Using the methods developed for Pivot Based SMT (Wu and Wang, 2007) (Utiyama and Isahara, 2007) one can use the source-pivot and …

Content truncated to 500 words