Statistical Post-Editing of Machine Translation for Domain Adaptation

Statistical Post-Editing of Machine Translation for Domain Adaptation Rapha¨el Rubino St´ephane Huet LIA-CERI Universit´e d’Avignon et des Pays de Vaucluse Avignon, France {firstname.lastname}@univ-avignon.fr Fabrice Lef`evre Georges Linar`es Abstract This paper presents a statistical approach to adapt out-of-domain machine trans- lation systems to the medical domain through an unsupervised post-editing step. A statistical post-editing model is built on statistical machine translation (SMT) out- puts aligned with their translation refer- ences. Evaluations carried out to trans- late medical texts from French to English show that an out-of-domain machine trans- lation system can be adapted a posteri- ori to a specific domain. Two SMT sys- tems are studied: a state-of-the-art phrase- based implementation and an online pub- licly available system. Our experiments also indicate that selecting sentences for post-editing leads to significant improve- ments of translation quality and that more gains are still possible with respect to an oracle measure. 1 Introduction Phrase-Based Machine Translation (PBMT) is a popular approach to Statistical Machine Transla- tion (SMT) that leads to accurate translation re- sults (Zens et al., 2002; Marcu and Wong, 2002; Koehn et al., 2003). The statistical models used in PBMT are based on the probabilities of bidirec- tional alignment of phrases between two sentences in the translation relation. The linguistic resources used to estimate such probabilities are parallel cor- pora and the main resulting statistical model is a translation table. Therefore, parallel corpora are the cornerstone for high quality translation. How- ever, such resources are expensive to construct. c⃝ 2012 European Association for Machine Translation. This lack of parallel data still remains an issue in PBMT. This phenomenon is accentuated by the diversity of texts to translate, in terms of origin and domain. As explained in (Sager et al., 1980), most of human activities involve a specific lan- guage or a subject language. A specific domain can be characterized by particular terminology or syntactic and discourse structures. As building do- main specific translation systems for each domain is unreasonable, we assume that domain adaptation of out-of-domain translation systems can be one of the solutions to address the diversity of specific do- mains. Although current machine translation systems can lead to impressive accuracy, translated texts re- quire sometimes human post-processing to be us- able. However, editing a posteriori can be costly depending on the amount of corrections required by machine translation outputs. Therefore, the automation of post-editing is an important task which can lead to higher quality machine transla- tion without requiring human intervention. In this paper, we propose a statistical post- editing (SPE) approach to adapt SMT systems to specific domains. We focus on translating texts in the medical domain from French to English. Sev- eral SMT systems are studied and we propose dif- ferent methods to include the in-domain data into the translation process. We evaluate how transla- tion quality can be improved with a post-editing step based on a phrase-based alignment approach. Two sets of experiments are presented in this pa- per: one applying SPE consistently on all the sen- …

Content truncated to 500 words