Pivot-Based Semantic Splicing for Neural Machine Translation

Current neural machine translation (NMT) usually extracts a fixedlength semantic representation for source sentence, and then depends on this representation to generate corresponding target translation. In this paper, we proposed a pivot-based semantic splicing model (PBSSM) to obtain a semantic representation including more translation information for source sentence, thus improving the translation performance of NMT. The spliced semantic representation is derived from source languages of trilingual parallel corpus by the pivot-based NMT. Besides, the proposed PBSSM only depends on one source language to generate its semantic representation during the encoding process. We integrated it into the NMT architecture. Experiments on the English-Japanese translation task show that our model achieves a substantial improvement by up to 22.9% (3.74 BLEU) over the baseline.