Research And Implementation Of Task-Oriented Dialogue System Based On Pipeline Method

Task-oriented multi-turn dialogue system construction methods mainly include the pipeline method and the end2end methods.The traditional pipeline method has too many modules,and there is error propagation between modules,which is difficult to optimize globally.The end2end-based method is to characterize the information between modules as feature vectors,which solves the problem of error propagation,but due to the lack of independent control of the dialogue management module,the model lacks robustness.Aiming at the problems of the above two dialogue systems,this paper proposes a new task-oriented multi-turn dialogue system based on pipeline method.Partial modules are jointly trained to reduce the number of modules and reduce error accumulation and propagation.Furthermore,this paper improves the dialogue state tracker,effectively utilizes the predefined ontology knowledge and enhances the ability of named entity recognition.An end2end training model for dialogue decision-making and dialogue generation based on deep learning methods is proposed to improve dialogue decision-making capabilities.Specifically,it includes the following three tasks:(1)Propose a dialogue state tracking method based on a pre-defined ontology knowledge base for enhanced named entity recognition.This method reduces the error propagation between the modules by jointly training two modules.The named entity recognition pointer makes full use of the named entity knowledge of the predefined ontology knowledge base to improve the named entity recognition ability and semantic understanding ability of the dialogue state tracker.And through the parameter sharing method,joint training of dialogue state inference pointers in multiple similar fields to improve the ability of dialogue state inference.This paper conducts experiments on the public data set MultiWOZ 2.1 of the task-oriented multi-turn dialogue system.The results show that the method proposed in this paper improves the dialogue state tracking ability by 1.2%compared with the existing model.(2)An end2end method for dialogue decision-making and dialogue generation based on deep learning methods is proposed.This method constructs the dialogue decision and dialogue generation module as an end2end module,which reduces the error propagation between modules.Through the representation learning and attention mechanism,it replaces the traditional construction of MYSQL query statements to query database information,which reduces the standardization process and improves the robustness of the model.This paper conducts experiments on the MultiWOZ 2.1 data set.The experimental results show that under a certain threshold,the database query accuracy rate of this paper is about 98.4%.And the experimental results show that the ability of dialogue decision-making has been significantly improved,and the fluency of dialogue has been improved.On the test set,the matching rate of experimental entities based on the Greddy search strategy reached 81.2%,an increase of 3%,and the success rate of slot mentions increased by 4.90%.Under the Beam strategy,the entity matching rate increased by 24.5%,the slot mention success rate increased by 12.90%,and the BLEU value increased by 2%.(3)Based on the research of the above two task-oriented multi-turn dialogue system modules,this paper builds a multi-domain task-oriented multi-turn dialogue system demonstration platform based on Flask's Web framework.Users can conduct restaurant inquiries and Hotel reservations,train ticket reservations,etc.complete different tasks in multiple fields.