Research On Task-oriented Multi-turn Dialogue Technology

In recent years,with the rapid changes and development of artificial intelligence and continuous breakthroughs in natural language understanding technology,the human-machine dialogue system has attracted the attention of a large number of researchers due to its broad application scenarios and great commercial value.Thanks to the rapid development of the Internet and the production of massive amounts of data,deep learning technology has made considerable progress,and end-to-end dialogue systems that rely on deep learning have gradually become the mainstream.The dialogue system based on the end-to-end method can directly generate corresponding output according to the input,which reduces the workload of manually labeling data and eliminates the error accumulation problem in the traditional pipeline method.However,the end-to-end method still has some problems that need to be solved urgently.For example,the generated answers may be very general or meaningless;although there is a memory function,the amount of memory is limited,and the memory time is limited.Processing long sequences is very time-consuming and difficult to apply to multiple applications.Round dialogue.This thesis proposes two models to solve the above problems,and implements a task-oriented multi-turn dialogue system based on the proposed model design.The specific work content can be summarized as follows:1.In response to the meaningless question of the generated answers,this article proposes a task-based dialogue model with a hybrid retrieval generation method,which combines retrieval models and sequence-to-sequence generation models.According to different situations,different model outputs are selected as responses.It retains the advantages of fast retrieval mode and high response consistency of the generation mode,and solves the problem that the retrieval model cannot handle input that is not strongly related to the knowledge base and the generation model may produce meaningless responses.2.Aiming at the memory storage problem,this article combines the memory network with the sequence-to-sequence model,uses the encoder-decoder structure,uses the memory network as the encoder,and the memory network plus the gate loop unit as the decoder.Combined with the memory network,the context can be saved.The characteristics of the sentence information,and the advantage that the sequence model can generate another sequence of different lengths according to one sequence,realizes a sequence model combined with a memory network to improve the model's memory ability,effectively integrate the external knowledge base information,and improve the traditional sequence to The problem of the limited memory of the sequence model,thereby improving the ability of multiple rounds of dialogue.3.Design and implement a task-oriented multi-turn dialogue system.The model proposed in this thesis is constructed as a dialogue engine,and a complete dialogue system is realized by combining the server and the web front end.The Web front-end receives user questions and returns the answers to the user;the server transmits the information between the front-end and the dialogue engine;the dialogue engine uses the trained dialogue model to generate a suitable response after receiving the user's input and returns it.Finally,through the test and evaluation of the system,the validity and feasibility of the model proposed in this thesis are verified.This system has been delivered to Xin Wang Bank and is in the testing phase,and will soon be put into use in Xin Wang Bank's customer service system.