Using Comparable Corpora to Adapt MT Models to New Domains

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 437–444, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Using Comparable Corpora to Adapt MT Models to New Domains Ann Irvine Center for Language and Speech Processing Johns Hopkins University Chris Callison-Burch Computer and Information Science Dept. University of Pennsylvania Abstract In previous work we showed that when us- ing an SMT model trained on old-domain data to translate text in a new-domain, most errors are due to unseen source words, unseen target translations, and in- accurate translation model scores (Irvine et al., 2013a). In this work, we target er- rors due to inaccurate translation model scores using new-domain comparable cor- pora, which we mine from Wikipedia. We assume that we have access to a large old- domain parallel training corpus but only enough new-domain parallel data to tune model parameters and do evaluation. We use the new-domain comparable corpora to estimate additional feature scores over the phrase pairs in our baseline models. Augmenting models with the new features improves the quality of machine transla- tions in the medical and science domains by up to 1.3 BLEU points over very strong baselines trained on the 150 million word Canadian Hansard dataset. 1 Introduction Domain adaptation for machine translation is known to be a challenging research problem that has substantial real-world application. In this set- ting, we have access to training data in some old- domain of text but very little or no training data in the domain of the text that we wish to translate. For example, we may have a large corpus of par- allel newswire training data but no training data in the medical domain, resulting in low quality trans- lations at test time due to the mismatch. In Irvine et al. (2013a), we introduced a tax- onomy for classifying machine translation errors related to lexical choice. Our ‘S4’ taxonomy in- cludes seen, sense, score, and search errors. Seen errors result when a source language word or phrase in the test set was not observed at all during training. Sense errors occur when the source lan- guage word or phrase was observed during train- ing but not with the correct target language trans- lation. If the source language word or phrase was observed with its correct translation during train- ing, but an incorrect alternative outweighs the cor- rect translation, then a score error has occurred. Search errors are due to pruning in beam search decoding. We measured the impact of each error type in a domain adaptation setting and concluded that seen and sense errors are the most frequent but that there is also room for improving errors due to inaccurate translation model scores (Irvine et al., 2013a). In this work, we target score errors, using comparable corpora to reduce their frequency in a domain adaptation setting. We assume the setting where we have an old- domain parallel training corpus but no new domain training corpus.1 We do, however, have access to a mixed-domain comparable …

Content truncated to 500 words