The Research Of Multi-Turn Dialogue Model Based On Deep Learning

As an important component of Human-robot interaction,verbal-text interaction is of key importance to achieving effective Human-robot communication.In the traditional single-turn dialogue model,the robot’s one-question-and-answer,unconnected and rigid responses will lead to a reduced dialogue experience.In this study,we focus on the use of deep learning technology to build the basic model of multi-talk.It is mainly transformed into a modular construction combining multi-turn single-utterance generation,multi-turn multi-utterance generation,and intention recognition context construction model to improve the dialogue fluency;and a multi-turn dialogue system is built by combining distributed technology and humanoid robot platform SHFR-III.The main research contents and innovations of this paper are as follows:(1)The multi-turn dialogue utterance generation model is studied.To address the difficulty of encoding the historical context in the process of multi-turn dialogue generation,MPBG and 3AMTG multi-turn single utterance generation models are proposed based on two fusion methods of representation layer and semantic layer,respectively.The experimental results show that the two models proposed in this paper outperform other models in terms of accuracy,BLEU,CHRF++,and other evaluation criteria based on the "Chinese Multi-turn dialogue dataset" open-sourced by Tsinghua University and Tencent.In addition,to exclude the chance interference,this paper also improves the pre-trained ELMO model and proposes the 3EATG model combined with the Transformer model to build a comparison experiment.The experimental results show that the semantic layer fusion model 3AMGT proposed in this paper outperforms other models in all three evaluation metrics.(2)To solve the problem of multi-utterance intention recognition and context construction in the process of continuous dialogue,two models,Con BTM and Multi-ELAM,are proposed in this paper.This model solves the problem that the traditional Markov chain-based recurrent neural network and other models lack the ability to fuse multiple sequences of information.Based on the Chinese multi-turn dialogue dataset,the two proposed models perform well in terms of accuracy,recall,and Micro-F1 score,and the Multi-ELAM model achieves the best recognition results,thus verifying the superiority of the models.(3)A multi-turn dialogue Human-robot interaction system is constructed.The multi-model integration of the single-turn dialogue generation model Pro BG,the multi-turn single-utterance generation model 3AMTG,and the continuous dialogue intention recognition and context construction model Multi-ELAM model.It is also combined with the humanoid robot platform SHFR-III to build the multi-turn conversational Human-robot interaction system.,which generates dialogue utterances through the language generation model,and constructs dialogue contexts and models through the intention recognition and context construction model scheduling to achieve continuous dialogue.And the overall language effect of the multi-talk model and the fluency of the dialogue process was also verified through human-robot dialogue experiments.