Design And Implementation Of Topic Modeling In Retrieval Dialogue System

In recent years,open domain dialog system(chatbot)has received more and more attention.Compared with traditional information service forms(such as question answering system,search engine,etc.),open domain dialog system provides richer semantic content and more effective interaction mode.These characteristics make it more and more popular in the foreseeable future life.Therefore,the research on related technologies of open domain dialog system It has great economic effect and social value.With the rapid development of deep learning in the field of natural language processing,the dialogue system is becoming mature with the development of portable mobile device market,which is widely used in customer service,finance,medical care,education and life services.Open domain dialog system can be divided into two ways according to the way of its reply generation: retrieval and generative:retrieval dialog system searches and sorts the words matching the user's query from the system's reply database through matching technology,and selects the reply with the highest ranking.This method has the characteristics of high reply quality,but it depends on a large number of manually written dialogue data;Generative dialogue system based on data-driven methods can train the reply generation model.This method has better flexibility and can automatically generate new replies,but it is easy to generate a general no content reply.At present,most of the research work on chat conversation model is based on the neural network method.An end-to-end neural network model generates conversation replies or sorts the candidate replies.The disadvantage of this end-to-end model is that it is difficult to explain the generation or ranking process.In order to increase the interpretability of the dialogue process and avoid the main problems of the two methods of open domain dialogue system,this paper studies the retrieval dialogue system that integrates the utterance's topic,predicts the key words and topic information of the reply through the current input and historical dialogue information of the user,and sorts the candidate replies by the key words and theme retrieval,so as to ensure the diversity of the reply.This paper mainly carries out the following three research work:(1)Key words extraction modeling of dialogue data.In order to obtain the topic information of conversation from a large number of unlabeled conversation data,thispaper manually annotates the chat data,proposes a keyword extraction model combining sequence annotation method and word segmentation information,extracts keywords from the conversation,and verifies the performance of the model through experiments.(2)Predictive modeling of dialogue topic and topic words.In order to model topic transfer and one to many phenomena in multi round conversation scenarios,a topic prediction model based on sequence generation method is proposed,and a VAE model is used to model keyword prediction tasks.Finally,the model is verified by experiments to achieve the expected results.(3)Build a retrieval dialogue system that integrates utterance's topics.Combined with the research content of the first two chapters,a retrieval dialogue system is proposed,which integrates the utterance's topic information,and uses the predicted reply topic and the keywords to retrieve and rank the candidate replies.And build a demonstration system,which can provide topic related reply to user's input.