The Research And Implementation Of Closed Domain Task-oriented Dialogue System Based On Deep Learning

With the rise of artificial intelligence,task-oriented dialogue system has always been one of the popular research directions of scholars.It is positioned as the entrance of future application services,which is used to replace the current Internet search engine and serve as a new traffic distribution portal,and has broad application prospects.Closed domain task-oriented dialogue system refers to information or services in a specific field where users have a clear purpose and want to meet certain restrictions.The requirements can be more complicated and need to be represented in multiple turns.For the real business scenarios in the limited field,this paper mainly studies the construction method of the closed domain task-oriented dialogue system based on the pipeline model.In the current public research work,there are still the following problems:(1)As for the dialogue understanding task,most of the user's input are short texts,which there are analytical difficulties such as grammar missing and colloquialization.There is still a large room for improvement in the accuracy of intent detection and slot filling.(2)As for dialogue management and generation,the technology based on probabilistic model or deep reinforcement learning has higher thresholds.The model becomes hard to explain and maintain.Also it is uncontrollable and difficult to apply.In view of the above problems,this paper focuses on the research and analysis of the key construction techniques of closed domain task-oriented dialogue system,and proposes a new model or scheme for its two components and carries out experimental verification.At the same time,it realizes the design and development of the dialogue system for medical registration information query.The specific work contents are as follows:(1)Research and implementation of joint modeling based on CNN-LSTMs in dialogue understanding,this paper proposes a model with semantic encoding of multi-class feature fusion and decoding of improved attention mechanism.The accuracy of intent detection is increased by 1.05%and the F1 value of slot filling is increased by 0.74%on the ATIS dataset.(2)Research and implementation of the dialogue management and generation modeling based on multi-classification model,this paper compares with the action prediction performance of various classification algorithms based on the actual medical scene dataset.The scheme combining the slot filling framework and GRU sequence network has better predicted performance;(3)We build an online demonstration platform for dialogue system for medical registration information query based on Flask and React framework,and apply the algorithm and scheme proposed in this paper to the actual medical business scenario for verification.