Research On Natural Language Dialogue Generation Methods Fusing Emotional Factor

Building dialogue systems that can naturally converse with human has been an important research goal of artificial intelligence.Natural language dialogue has a wide range of applications in customer service,online learning,medical system and other fields,which attracts a great deal of attention from more and more researchers.Existing works in natural language dialogue are mainly implemented by the sequence-to-sequence framework.However,these dialogue systems mainly focus on the semantic relevance of response generation,ignoring the effect of emotion state on the response generation.Emotion state plays a very important role in people's daily communication,and the rich emotion state helps the speaker to express his thought.Therefore,it is very necessary to consider the role of emotion factor in the natural language dialogue system.In addition,word repetition,word omission,and other grammatical errors sometimes occur in responses generated based on the sequence to sequence model.In this paper,in order to address these two problems,we propose the natural dialogue generation method based on emotion encoding,and the emotional dialogue generation method based on the hierarchical attention model.The main research contents of this thesis are as follows:1)We propose a method of natural dialogue generation based on emotion encoding.This method encodes the emotion state of the input sequence as a vector representation that is fused into the process of response generation,which is used to promote the emotional expression of generating responses.The first stage is encoding feature learning.The input sequence is semantically and emotionally encoded to obtain the encoding feature sequence and the emotion vector representation,respectively.The second stage is decoding feature learning.Decoding hidden feature is learned by combining the emotion representation,context vector and previous generation word into the decoder.The third stage is output sequence generation.The word prediction error and emotion loss are combined to jointly learn the sequence generation model by the multi-task learning mechanism,and a novel re-rank function is introduced to select the appropriate response from the generated response list.The experimental results on the STC dataset show that this method can effectively improve the emotion expression of the generated responses compared with the current sequence-to-sequence dialogue model,and achieves better results in automatic metrics and human evaluation.2)We propose a method of emotional dialogue generation based on the hierarchical attention model.The method uses the hierarchical attention model to dymatically calculate the contribution of the emotion and semantic to response generation.The main purpose is to further solve grammatical problems that occur when dialogue model generates responses integrating emotion state.The hierarchical attention model mainly includes two layers of structure.The first-layer is the word-level contribution calculation.At each moment,the attention mechanism is used to calculate the real-time attention contribution of the encoding feature sequence and the previous generated word,and uses the contribution as the corresponding weight to calculate the semantic context vector at the current moment.The second-layer accepts the semantic context vector and the emotion state vector as input,and adopts the attention mechanism to calculate the real-time attention contribution of the semantic context and the emotion state,then uses contribution as the weight of the corresponding input to calculate the new context vector.The context vector is finally input into the decoder for the emotional dialogue model learning.The results on the dialogue dataset STC show that the method effectively solves the grammatical problems that arise when dialogue model generates responses.3)Design and implement the prototype system for natural language dialogue fusing emotional factor.PyQt is used to design the user interface of the prototype system,and the core algorithm of the system is implemented using Python,Tensorflow and Numpy.The system includes two modules for dialogue model learning and dialogue model testing.The implementation of prototype system verifies the availability and effectiveness of the proposed methods.