Application And Research Of Chatbot Based On Retrieval And Generation Method

As an important branch of natural language processing,intelligent chatbots have been widely used in practical scenarios such as smart speakers,online customer service,and voice assistants,which have become one of the hottest and most challenging research directions.By analyzing and understanding human language,intelligent chatbots can not only output meaningful responses in daily communication but also assist humans in completing specific tasks.Therefore,chatbots have always haves huge commercial value and broad development space.The implementation methods of chitchat chatbots are mainly divided into two types:retrieval-based methods and generation-based methods.Given the conversation context,the retrieval-based method finds the best matched response from the conversation history retrieval database,and the generation-based method generates the response according to the conversation context.However,these two methods have their own advantages and disadvantages.The retrieval-based method is limited in the size of the retrieval database and lacks flexibility.The generation-based method is easy to output trivial responses without information.To cope with these issues,our thesis proposes a parallel joint method of chatbots by combining the advantages of both retrieval-based methods and generation-based methods.The chatbot comprehensively considers the replies generated by the two methods,and finally selects the best response of the two methods as the output.In the dialogue retrieval network,this thesis proposes and designs a retrieval model based on the deep residual matching mechanism,which is used to solve the problem of semantic drift in matching and greatly improve the reasoning speed of the model.In the dialogue generation network,our thesis uses a large-scale pre-training model based on post-training to improve the model’s ability to solve ordinary reply problems.In the final selection of candidate responses,our thesis proposes a dialogue response hybrid ranking network based on maximum mutual information,which helps the model to select the best response without annotated data training.The comparison and analysis of the experimental results show that the chitchat chatbot based on the parallel joint method constructed in this thesis can output high fluency,strong relevance and information-rich responses.This thesis also deploys the algorithm to practical applications,and constructs a web-based chatbot.The overall structure of the chatbot is simple and easy to deploy,and can provide chat services at any time.