Integrating TM Knowledge into NMT with Double Chain Graph

Abstract

The approach based on translation pieces extracted from the translation memory (TM) knowledge is appealing for neural machine translation (NMT), owning to its efficiency in memory consumption and computation. However, the incapable of capturing sufficient contextual translation knowledge leading to a limited translation performance. This paper proposes a simple and effective structure to address this issue. The main idea is to employ the word chain and position chain knowledge from a TM as additional rewards to guide the decoding process of the neural machine translation. Experiments on six translation tasks show that the proposed Double Chain Graph yields consistent gains while achieving greater efficiency to the counterpart of translation pieces.