Hierarchy Response Learning for Neural Conversation Generation

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1772–1781, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1772 Hierarchy Response Learning for Neural Conversation Generation Bo Zhang1 and Xiaoming Zhang1,2 1 School of Cyber Science and Technology 2 Hefei Innovation Research Institute Beihang University, China {zhangnet,yolixs}@buaa.edu.cn Abstract The neural encoder-decoder models have shown great promise in neural conversation generation. However, they cannot perceive and express the intention effectively, and hence often generate dull and generic re- sponses. Unlike past work that has focused on diversifying the output at word-level or discourse-level with a flat model to alleviate this problem, we propose a hierarchical gen- eration model to capture the different levels of diversity using the conditional variational autoencoders. Specifically, a hierarchical re- sponse generation (HRG) framework is pro- posed to capture the conversation intention in a natural and coherent way. It has two modules, namely, an expression reconstruction model to capture the hierarchical correlation between expression and intention, and an expression at- tention model to effectively combine the ex- pressions with contents. Finally, the training procedure of HRG is improved by introducing reconstruction loss. Experiment results show that our model can generate the responses with more appropriate content and expression. 1 Introduction Neural conversation generation (Xu et al., 2017; Zhang et al., 2018), focusing on responding to hu- mans intelligently on a variety of topics, has drawn great attention from both academic and indus- try. The sequence-to-sequence model (Seq2Seq) (Sutskever et al., 2014) is one type of neural gen- eration model that maximizes the probability of generating a response given the dialogue context. It enables the incorporation of rich context to gen- erate coherent responses in an unsupervised man- ner. However, it was found that Seq2Seq models suffer from so-called safe response problem (Xu et al., 2017) , i.e., they tend to generate some dull and generic repetitive responses (e.g., “I think so”, “I don’t know’, etc.), rather than meaningful and conscious expression. Xu et al. (2017) ascribed this to the fundamental nature of statistical models since the distribution of most pieces of informa- tion are relatively sparser when compared to the safe response patterns in the open domain conver- sations. Some works attempted to improve the ar- chitecture of Seq2Seq models, including introduc- ing reinforcement learning (Zhang et al., 2018), encouraging responses that have long-term pay- off, etc. The other important reason is that the response generation model cannot express the in- tention and emotion internally. Thus, one line of research has focused on forcing the model to simu- late some human’s skills by augmenting the input with rich meta information. For example, some recent works biased the responses to some spe- cific personas (Li et al., 2016b) or emotions (Hu- ber et al., 2018). Usually, in the process of human conversation, a speaker participates in the dialogue including the following steps. The speaker is firstly required to decide what the …

Content truncated to 500 words