A Neural Conversational Model Using MMI-WMD Decoder Based on the Seq2Seq with Attention Mechanism
Tianzhi Wang, Meng Cai, Jianxun Li · 2019
Sequence-to-sequence neural conversational model generates responses tend to be safe and common-place(e.g., I don't know, I'm not sure, I'm sorry), meanwhile the more diversity-promoting and conform responses to queries are not scored higher probably in the N-best lists. We suggest the Maximum Mutual Information(MMI) for diversifying the responses combing with Word Mover's Distances(WMD) for relating the responses to queries as the objective function to rerank N-best lists based on the Seq2Seq with attention mechanism. Experimental results prove that the MMI-WMD model generates more diverse, meaningful, relevant and appropriate responses to queries. The model performs well in human evaluations and BLEU scores in cornell movie dialogue datasets.