Neural Dialogue Model with Retrieval Attention for Personalized Response Generation.

With the success of new speech-based human-computer interfaces, there is a great need for effective and friendly dialogue agents that can communicate with people naturally and continuously. However, the lack of personality and consistency is one of critical problems in neural dialogue systems. In this paper, we aim to generate consistent response with fixed profile and background information for building a realistic dialogue system. Based on the encoder-decoder model, we propose a retrieval mechanism to deliver natural and fluent response with proper information from a profile database. Moreover, in order to improve the efficiency of training the dataset related to profile information, we adopt a method of pre-training and adjustment for general dataset and profile dataset. Our model is trained by social dialogue data from Weibo. According to both automatic and human evaluation metrics, the proposed model significantly outperforms standard encoder-decoder model and other improved models on providing the correct profile and high-quality responses. [ABSTRACT FROM AUTHOR]