Question Rewrite Based Dialogue Response Generation

摘要

Dialogue response generation is a fundamental technique in natural language processing, which can be used in human-computer interaction. As the quick development in neural networks, the sequence to sequence (seq2seq) model which employed recurrent neural networks (RNN) encoder-decoder has archived great success in machine translation. Many researchers began to apply this model in dialogue response generation. However, the conventional seq2seq model counters several problems, e.g., grammatical mistake, safe response and etc. In this paper, motivated by the great success of generative adversarial networks (GANs) in generating images, we propose an improved seq2seq framework by employing GANs to rewrite questions in order to retrieve more information from the question. Afterwards we combine the original question and the rewritten question together to generate responses. The experiments on the public Yahoo! Answers dataset demonstrated the proposed framework's potential in dialogue response generation.