Research And Implementation Of Emotional Conversation Generation System Based On Deep Learning

With the continuous development of deep learning,natural language processing technology is developing rapidly and is gradually applied to various fields,such as semantic analysis,text sentiment analysis,machine translation,question answering system,etc.Among them,the task of natural language dialogue generation makes use of natural language processing technology based on deep learning,which can generate diverse and creative responses.The goal of this task is to achieve the interaction between machines and humans,and try to model the interacting process as close as possible to the human-level conversation.This thesis proposes an emotional dialogue generation system based on deep learning,which can learn and model the emotional expression in the dialogue,and obtains the capability of adaptive emotional perception and expression in open-domain dialogue scenarios.The main contributions and innovations of this thesis are as follows:First of all,in order to guarantee the quality of generated language,this thesis starts with the reseach of traditional dialogue genenration,and conducts in-depth analysis and optimization of the problems in the research of response diversity and dialogue relevance.Specifically,to improve the diversity of responses,this thesis uses cross-entropy loss function with inverse token frequency weight restriction to alleviate the problem of dull and universal response generation,which is often caused by the previous training objective optimization based on maximum likelihood estimation.Furthermore,to improve the relevance of generated dialogues,this thesis proposes a bi-directional generation framework based on dual optimization and training,which uses two Sequence-to-Sequence models to jointly learn the information association from dialogue contexts in two directionsSecondly,on the basis of traditional dialogue generation,this thesis proposes an emotional dialogue generation model with the capability of adaptive emotion controlment and expression,which achieves the modeling of emotional expression in conversation in two levels.First,the emotion category controlling module in the model makes control over the response emotion types.By learning and using contextual emotion mapping relationship in dialogue,it selects a reasonable emotion category for response during the conversation.Then,based on the determined emotion category,the emotional decoder can generate target response in fine-grained emotion level,which achieves the potential emotion state decay in dialogue and uses relevant emotion words selectively to generate fluent and emotional responses.In addition,this thesis also proposes the corresponding construction strategy of emotion distribution labels and emotion word vectors in emotion learning.Finally,this thesis uses the above model algorithms and data to implement an open-domain daily chitchat system,which is presented on the web page to achieve emotional interaction with users effectively.It can flexibly deal with chitchat topics and gives user-friendly experience.