A novel method to optimize training data for translation model adaptation

In this paper, we explore the method to improve the cross-domain adaptation of current translation models, with the aim to solve the common problem that ambiguous linguistic knowledge in different domain causes a difficult training for a robust translation model. Specially, we propose a novel method to automatically optimize training data for translation model adaptation. The method combines a test sentence and its best candidate translation to generate a pseudo-parallel translation pair. Regarding the pairs as queries, the method follows a twin-track retrieval approach to further mine parallel sentence pairs from large-scale bilingual resources. Experiments show that by using our method, the optimized translation models significantly improve the translation performance by 1.8 BLEU points when only 7.7% of bilingual training data is used.