Combining Bilingual and Comparable Corpora for Low Resource Machine Translation

Proceedings of the Eighth Workshop on Statistical Machine Translation, pages 262–270, Sofia, Bulgaria, August 8-9, 2013 c⃝2013 Association for Computational Linguistics Combining Bilingual and Comparable Corpora for Low Resource Machine Translation Ann Irvine Center for Language and Speech Processing Johns Hopkins University Chris Callison-Burch∗ Computer and Information Science Dept. University of Pennsylvania Abstract Statistical machine translation (SMT) per- formance suffers when models are trained on only small amounts of parallel data. The learned models typically have both low accuracy (incorrect translations and feature scores) and low coverage (high out-of-vocabulary rates). In this work, we use an additional data resource, compa- rable corpora, to improve both. Begin- ning with a small bitext and correspond- ing phrase-based SMT model, we improve coverage by using bilingual lexicon induc- tion techniques to learn new translations from comparable corpora. Then, we sup- plement the model’s feature space with translation scores estimated over compa- rable corpora in order to improve accu- racy. We observe improvements between 0.5 and 1.7 BLEU translating Tamil, Tel- ugu, Bengali, Malayalam, Hindi, and Urdu into English. 1 Introduction Standard statistical machine translation (SMT) models (Koehn et al., 2003) are trained using large, sentence-aligned parallel corpora. Unfortu- nately, parallel corpora are not always available in large enough quantities to train robust models (Ko- lachina et al., 2012). In this work, we consider the situation in which we have access to only a small amount of bitext for a given low resource language pair, and we wish to supplement an SMT model with additional translations and features estimated using comparable corpora in the source and tar- get languages. Assuming access to a small amount ∗Performed while faculty at Johns Hopkins University of parallel text is realistic, especially considering the recent success of crowdsourcing translations (Zaidan and Callison-Burch, 2011; Ambati, 2011; Post et al., 2012). We frame the shortcomings of SMT models trained on limited amounts of parallel text1 in terms of accuracy and coverage. In this con- text, coverage refers to the number of words and phrases that a model has any knowledge of at all, and it is low when the training text is small, which results in a high out-of-vocabulary (OOV) rate. Accuracy refers to the correctness of the transla- tion pairs and their corresponding probability fea- tures that make up the translation model. Because the quality of unsupervised automatic word align- ments correlates with the amount of available par- allel text and alignment errors result in errors in extracted translation pairs, accuracy tends to be low in low resource settings. Additionally, esti- mating translation probabilities2 over sparse train- ing sets results in inaccurate feature scores. Given these deficiencies, we begin with a base- line SMT model learned from a small parallel cor- pus and supplement the model to improve its ac- curacy and coverage. We apply techniques pre- sented in prior work that use comparable corpora to estimate similarities between word and phrases. In particular, we build on prior work in bilingual lexicon induction in order to predict translations for …

Content truncated to 500 words