Efficient Diverse Response Generation in Attention-based Neural Conversational Model with Maximum Mutual Information

Diversity} of generated responses is important for a data-driven neural conversational model (NCM) for non-task-oriented conversation. A criterion of maximum mutual information (MMI) and generating N-best outputs are both effective ways to increase the diversity. Generally, a beam search is used for generating N-best outputs. However, the beam search is likely to produce similar outputs in the N-best results. We propose a simple and efficient N-best search, namely N-greedy search, for an encoder-decoder recurrent neural network (RNN) with an attention mechanism. We built an NCM with a fictive chitchat corpus and generated responses based on the MMI criterion and N-greedy search. All of four objective indices of diversity showed increases, and a subjective evaluation clearly showed a reduction in the number of dull responses.