Domain Adaptation for Statistical Machine Translation with Domain Dictionary and Monolingual Corpora

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 993–1000 Manchester, August 2008 Domain Adaptation for Statistical Machine Translation with Domain Dictionary and Monolingual Corpora Hua Wu, Haifeng Wang Toshiba (China) R&D Center Beijing, 100738, China wuhua@rdc.toshiba.com.cn wanghaifeng@rdc.toshiba.com.cn Chengqing Zong NLPR, Institute of Automation Chinese Academy of Sciences Beijing 100080, China cqzong@nlpr.ia.ac.cn Abstract tra Statistical machine translation systems are usually trained on large amounts of bilingual text and monolingual text. In this paper, we propose a method to per- form domain adaptation for statistical machine translation, where in-domain bi- lingual corpora do not exist. This method first uses out-of-domain corpora to train a baseline system and then uses in-domain translation dictionaries and in-domain monolingual corpora to improve the in- domain performance. We propose an al- gorithm to combine these different re- sources in a unified framework. Experi- mental results indicate that our method achieves absolute improvements of 8.16 and 3.36 BLEU scores on Chinese to English translation and English to French translation respectively, as compared with the baselines using only out-of- domain corpora. 1 Introduction In statistical machine translation (SMT), the translation process is modeled to obtain the translation of the source sentence f by maximizing the following posterior probability (Brown et al., 1993). best e () ) ( max arg ) ( max arg | | e e ef e f e e LM best p p p = = (1) State-of-the-art SMT systems are trained on large collections of bilingual corpora for the ○ C 2008. Licensed under the Creative Commons Attri- bution-Noncommercial-Share Alike 3.0 Unported license (http://creativecommons.org/licenses/by-nc- sa/3.0/). Some rights reserved. nslation model pf( | e) and monolingual tar- get language corpora for the language model (LM) pLM(e) . The trained SMT systems are suitable for translating texts in the same domain as the training corpus. However, for some spe- cific domains, it is difficult to obtain a bilingual corpus. In this case, the performance of SMT systems will be degraded. Generally, it is easier to obtain in-domain monolingual corpora in either source or target language. Moreover, in some specific domains, although in-domain bilingual corpora do not ex- ist, in-domain translation dictionaries, which usually contain domain-specific terms and their translations, are available. And even if such dic- tionaries are not available, it is easier to compile one than to build a bilingual corpus. Thus, in this paper, we address the problem of domain- specific SMT, where only domain-specific dic- tionaries and/or monolingual corpora exist. In a specific domain, there are two kinds of words: common words, which also frequently occur in out-of-domain corpora, and domain-specific words, which only occur in the specific domain. Thus, we can combine the out-of-domain bilin- gual corpus, the in-domain translation dictionary, and monolingual corpora for in-domain transla- tion. If an in-domain translation dictionary is avail- able, we combine it with the out-of-domain translation model to improve translation quality. If an in-domain target language corpus (TLC) is available, we use it to build an in-domain lan- guage model, which can …

Content truncated to 500 words