Multi-turn Dialogue System Based on Improved Seq2Seq Model

Abstract

The automatic dialogue system is an intelligent system built by combining various artificial intelligence technologies. In recent years, with the introduction of multi-turn dialogue generation systems, semantic relevance and topic consistency between different dialogue turns have become important evaluation criteria for the success of the model. However, these problems have not been resolved and still face many challenges. In this paper, we propose a sequence to sequence (seq2seq) model based on multi-encoder structure and theme-oriented decoder, and these innovations enable the proposed model to obtain semantic correlation between different turns of conversation and maintain the consistency of topics. Meanwhile, aiming at the disadvantages of the basic seq2seq model, BiLSTM cells, attention mechanism and beam search algorithm are adopted to solve the problem of long-distance dependence and obtain richer semantic information. Experiments show that the proposed model can generate coherent, appropriate and diverse replies on multi-turn dialogue datasets.