关键词: 对话生成, 深度增强学习, 分层编码, 循环神经网络, 序列到序列

Abstract: Aimed at dialog generation problem, a dialog generation model is proposed to solve the problem that standard seq2seq architectures are more likely to raise highly generic responses due to the Maximum Likelihood Estimate(MLE)loss function. This method combines hierarchical encoding method and deep reinforcement learning, and uses hierarchical structure to build a multi-round dialog model, adding a hierarchical layer to enhance the memory of history dialog based on the standard seq2seq architecture, and then use a language model to build reward functions, replacing traditional MLE loss function with policy gradient method in deep reinforcement learning for training. Experimental results show that the proposed model (Enhanced HRED) can generate responses with richer semantic information and improve by 17%-21% in standard manual evaluation compared with widely used traditional RNN dialog generation models.

Key words: dialog generation, deep reinforcement learning, hierarchical encoding, recurrent neural network, sequence to sequence(seq2seq)