A Bichannel Transformer with Context Encoding for Document-Driven Conversation Generation in Social Media

Table 5

Ablation study on document relevance and response quality.

Models	Doc_BLEU	NW	avg_len	PPL
BCTCE	28.23	0.36	9.16	17.80
+copy	27.38	0.34	8.63	16.33
-document	26.25	0.23	8.42	12.97
-history	28.47	0.65	8.05	15.97
-context encoder	23.66	0.39	6.72	11.23
-utterance encoder	28.25	0.41	9.40	16.88
-bi_channel	29.15	0.33	7.94	14.26