Research On Intelligent Dialogue Q&A System Based On Deep Learning

Intelligent question-answering system has long been a research focus in the field of natural language processing.As the technique of deep learning has been gradually getting mature in recent years,more and more researches adopt the deep neural network for text semantic modeling.Deep neural network is a compute model ed through simulating a human brain cognitive process,which is closer to the cognitive mode of human being.In this paper,we use the deep neural network method to focus on the generation of dialogue generation problem,the problem of question matching based on semantic comprehension and the sentence classification based on semantic representation learning.The main research contents include the following four aspects:(1)The man-machine dialogue research is based on seq2seq model and reinforcement learning.In order to solve the problems of dull reply and dialogue infinite loop in the dialogue task of seq2seq model,this thesis learn from(Jiwei Li,2016)idea that puts forward the objective function and introduces the reinforcement learning method to imitate dialogue practice,which make the reply more accurate and qualified.The experiment result shows that the improved seq2seq model avoid the dull replay and dialogue infinite loop to some extent.(2)Semantic comprehension has long been the research focus in the research of the question-answer system.The traditional research method is to analyze and comprehend the semantics of questions by the steps of segmentation,named entity recognition,POS Tagging,syntax parsing,keyword extraction.However,this thesis regards the questions as a sequence whole to do semantic computing,adopts the recurrent neural network model with bidirectional GRU to model the sequence of questions.The experiment result shows that this semantic comprehension question-answer system can make full use of the context to get semantic matching answer.(3)Since the system has the functions of entertainment,chatting and intelligent dialogues,This thesis adopts the recurrent convolutional neural network method to classify the messages.This method combines the advantages of semantic sequence modeling of the recurrent neural network and text feature extraction of the convolutional neural network method.The experiment result shows that this method can make better use of the semantic information and text feature to classify the messages than the traditional methods such as machine learning.(4)The intelligent question-answer system is applied to the Wechat public account to provide service,To avoid the system installation and platform design and other complicated steps,By following the Wechat public account Xiaoxia,clients have access to the service.To conclude,it is feasible to construct an intelligent question-answer system which can be used to entertain,chat,answer the questions about tourism in Xiamen and Xiamen University by adopting the technique of deep learning,and apply the system to Wechat public account,which is more useful than the traditional methods.