Domain adaptation for statistical machine translation in development corpus selection

The performance of statistical machine translation (SMT) system is affected by model parameters (e.g. weights of feature functions), which are usually tuned on a development corpus. Most research done to date has focused on algorithms for tuning parameters. However, the selection of development corpus is lack of discussion. It is believed that the parameters trained on a proper corpus will improve translation performance. Instead of exploring new algorithms, this paper aims to select development corpus for tuning parameters according to the test set. We address this problem as domain adaptation and propose two methods based on information retrieval (IR) technique and text clustering (TC) technique, respectively. Experimental results show that both the methods yield more stable performance for tuning parameters than subjective selection of development corpus.