Improving Neural Machine Translation by Retrieving Target Translation Template

In the neural machine translation (NMT) paradigm, transformer-based NMT has achieved great progress in recent years. It is based on the standard end-to-end structure, and acquires translation knowledge through the attention mechanism from the parallel corpus automatically without human intervention. Inspired by the process of translating sentences by human translators and the successful application of translation template in statistical machine translation, this paper proposes a novel approach to incorporate the target translation template into the Transformer-based NMT model. Firstly, the template extraction method derives the parallel templates corpus from the constituency parse tree. Secondly, given a sentence to be translated, a fuzzy matching strategy is proposed to calculate the most possible target translation template from the parallel template corpus. Finally, an effective method is proposed to incorporate the target translate template into the Transformer-based NMT model. Experimental results on three translation tasks demonstrate the effectiveness of the proposed approach and it improves the translation quality significantly.