Source Language Adaptation for Resource-Poor Machine Translation

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 286–296, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Source Language Adaptation for Resource-Poor Machine Translation Pidong Wang Department of Computer Science National University of Singapore 13 Computing Drive Singapore 117417 wangpd@comp.nus.edu.sg Preslav Nakov QCRI Qatar Foundation Tornado Tower, P.O. 5825 Doha, Qatar pnakov@qf.org.qa Hwee Tou Ng Department of Computer Science National University of Singapore 13 Computing Drive Singapore 117417 nght@comp.nus.edu.sg Abstract We propose a novel, language-independent approach for improving machine translation from a resource-poor language to X by adapt- ing a large bi-text for a related resource-rich language and X (the same target language). We assume a small bi-text for the resource- poor language to X pair, which we use to learn word-level and phrase-level paraphrases and cross-lingual morphological variants be- tween the resource-rich and the resource-poor language; we then adapt the former to get closer to the latter. Our experiments for Indonesian/Malay–English translation show that using the large adapted resource-rich bi- text yields 6.7 BLEU points of improvement over the unadapted one and 2.6 BLEU points over the original small bi-text. Moreover, combining the small bi-text with the adapted bi-text outperforms the corresponding com- binations with the unadapted bi-text by 1.5– 3 BLEU points. We also demonstrate applica- bility to other languages and domains. 1 Introduction Statistical machine translation (SMT) systems learn how to translate from large sentence-aligned bilin- gual corpora of human-generated translations, called bi-texts. Unfortunately, collecting sufficiently large, high-quality bi-texts is hard, and thus most of the 6,500+ world languages remain resource-poor. For- tunately, many of these resource-poor languages are related to some resource-rich language, with whom they overlap in vocabulary and share cog- nates, which offers opportunities for bi-text reuse. Example pairs of such resource rich–poor lan- guages include Spanish–Catalan, Finnish–Estonian, Swedish–Norwegian, Russian–Ukrainian, Irish– Gaelic Scottish, Standard German–Swiss Ger- man, Modern Standard Arabic–Dialectical Arabic (e.g., Gulf, Egyptian), Turkish–Azerbaijani, etc. Previous work has already demonstrated the ben- efits of using a bi-text for a related resource-rich language to X (e.g., X=English) to improve ma- chine translation from a resource-poor language to X (Nakov and Ng, 2009; Nakov and Ng, 2012). Here we take a different, orthogonal approach: we adapt the resource-rich language to get closer to the resource-poor one. We assume a small bi-text for the resource-poor language, which we use to learn word-level and phrase-level paraphrases and cross-lingual morpho- logical variants between the two languages. Assum- ing translation into the same target language X, we adapt (the source side of) a large training bi-text for a related resource-rich language and X. Training on the adapted large bi-text yields very significant improvements in translation quality com- pared to both (a) training on the unadapted version, and (b) training on the small bi-text for the resource- poor language. We further achieve very sizable im- provements when combining the small bi-text with the large adapted bi-text, compared to combining the former with the unadapted bi-text. While …

Content truncated to 500 words