Human–Machine Multi-Turn Language Dialogue Interaction Based on Deep Learning

Xianxin Ke, Ping Hu, Chenghao Yang, Renbao Zhang

During multi-turn dialogue, with the increase in dialogue turns, the difficulty of intention recognition and the generation of the following sentence reply become more and more difficult. This paper mainly optimizes the context information extraction ability of the Seq2Seq Encoder in multi-turn dialogue modeling. We fuse the historical dialogue information and the current input statement information in the encoder to capture the context dialogue information better. Therefore, we propose a BERT-based fusion encoder ProBERT-To-GUR (PBTG) and an enhanced ELMO model 3-ELMO-Attention-GRU (3EAG). The two models mainly enhance the contextual information extraction capability of multi-turn dialogue. To verify the effectiveness of the two proposed models, we demonstrate the effectiveness of our model by combining data based on the LCCC-large multi-turn dialogue dataset and the Naturalconv multi-turn dataset. The experimental comparison results show that, in the multi-turn dialogue experiments of the open domain and fixed topic, the two Seq2Seq coding models proposed are significantly improved compared with the current state-of-the-art models. For specified topic multi-turn dialogue, the 3EAG model has the average BLEU value reaches the optimal 32.4, which achieves the best language generation effect, and the BLEU value in the actual dialogue verification experiment also surpasses 31.8. for open-domain multi-turn dialogue. The average BLEU value of the PBTG model reaches 31.8, the optimal 31.8 achieves the best language generation effect, and the BLEU value in the actual dialogue verification experiment surpasses 31.2. So, the 3EAG model is more suitable for fixed-topic multi-turn dialogues for the two tasks. The PBTG model is more muscular in open-domain multi-turn dialogue tasks; therefore, our model is significant for promoting multi-turn dialogue research.