Abstract

In this paper, we introduce context informed features in a log-linear phrase-based SMT framework; these features enable us to exploit source similarity in addition to target similarity modeled by the language model. We present a memory-based classification framework that enables the estimation of these features while avoiding sparseness problems. We evaluate the performance of our approach on Italian-to-English and Chinese-to-English translation tasks using a state-of-the-art phrase-based SMT system, and report significant improvements for both BLEU and NIST scores when adding the context-informed features.