Learning a Log-Linear Model with Bilingual Phrase-Pair Features for Statistical Machine Translation

We propose a set of informative feature functions togheter with a log-linear model framework for bilingual phrase-pair extraction to improve phrase-based statistical machine translation. The base feature functions investigated are phrase length model, phrase-level centers' distortion, lexicon translation equivalence, bracketing constraints and word alignment links. Two generative models show strong baselines withe these base features, illustrating the effectiveness of the proposed feature functions. Strategies of extending the features and a log-linear model of them are proposed to effectively extract phrase-pars from parallel data. Experimental results of TIDES'03 Chinese-English small data track show improved translation qualities.