Research And Implementation Of Task-oriented Dialogue System For Multi-domains

Task-oriented dialogue system on deep learning model has been widely concerned because of its broad application scenarios and rich technical challenges.More and more complex scenarios also lead to the emergence of multi domain task-oriented dialogue system.For example,when intelligent assistant helps to book tickets,it also needs to pay attention to the accommodation situation of the destination.The method of single-domain superposition will have problems such as slot sharing difficulty and system call chaos,so it is difficult to be applied to complex multi-domain scenarios.In multi-domain area,current research still has the problem of low accuracy in DST.The current mainstream methods have obvious limitations under the increasingly large number of ontologies,complex slot settings and other multi domain dialogue state tracking features.The number of dialogue actions and the long-distance dependence of dialogue history brought by multi-domain dialogue are also lacking.In view of the above problems faced by the multi domain dialogue system,this paper studies and implements a multi domain task driven dialogue system.The main work of this paper includes the following three points:(1)This paper proposes a self-attention dialogue state generation model based on joint learning.The model combines the traditional natural language understanding module and the DST module to avoid the error propagation between the two modules.We use the generation task to replace the sequence annotation task while combines slot classification and generation task by cross-attention and joint learning.The method above solves the problem of inaccurate slot and historical information loss.Out model achieves 54.18%of joint accuracy and 97.83%of slot accuracy on MultiWOZ which is better than other generation-based models.(2)A Turn-level self-attention dialogue history encoder is proposed.DAP and dialogue generation tasks are jointly studied and turn-level self-attention is designed.This refers to the cognitive logic of human beings for historical information when chatting,so as to avoid paying attention to unreasonable subsequent turns of information.At the same time,we use graph structured dialogue action representation which is expressed at the level of model structure.It enhances the efficiency of information interaction between multiple fields.On MultiWOZ,our proposed model achieves 72.3%of dialogue success rate.(3)On the basis of the above improvements,this paper builds a complete multi-domain dialogue system display platform,through the design and completion of front interface,database management,background control,session management and other sub services,realizes user-friendly interaction,account management,task management and other functions.Compared with single domain dialogue,the platform has more practical application value.