Improving Neural Machine Translation with AMR Semantic Graphs

Table 4

Experimental results on ConvSeq2Seq and Transformer.

Model	BLEU
	tst2013	tst2015
ConvSeq2Seq	26.98	24.78
ConvSeq2Seq-AMR	27.30	24.70
ConvSeq2Seq-AMR-F	27.40	25.20
ConvSeq2Seq-AMR-B	26.73	24.53
Transformer	28.60	25.76
Transformer-AMR	28.69	25.91
Transformer-AMR-F	28.88	26.28
Transformer-AMR-B	28.69	26.01

The bold values are the highest results for each group of models.