Topic-Net Conversation Model.

Most sequence-to-sequence neural conversation models have a ubiquitous problem that they tend to generate boring and safe responses with almost none useful information, such as "I don't know" or "I'm OK". In this paper, we study the response generation problem and propose a topic-net conversation model (TNCM) via incorporating topic information into the sequence-to-sequence framework. TNCM generates every response word using not only its word embedding hidden state but also the topic embedding, which guides the model to form more interesting and informative responses in conversation. The model increases the possibility of the topic word appearing in the response further via a mixed probabilistic model of two modes, namely the generate-mode and the topic-mode. Moreover, we improve the process of beam search during the test, which enhances the performance with better efficiency. Evaluation results on large scale dataset indicate that our model can significantly outperform state-of-the-art methods on response generation of the conversation system.