Research On Task-oriented Dialogue Understanding Method Based On Domain Knowledge Graph

In the studying of artificial intelligence,the intelligent dialogue system,which simulates human communication ability,has been the focus of industry for its extensive application scenarios.It is also one of the hot research directions of domestic and foreign research scholars.The task-oriented dialogue system collects user demand information through multiple rounds dialogue,and provides services that meet the requirements of user description.The task-oriented dialogue system for vertical domain not only provides more efficient and high-quality service for people life,but also saves labor cost for enterprises,which has great commercial application value.The system based on pipeline is usually composed of three modules(natural language understanding,dialogue management and natural language generation)and a back-end database.The accuracy of natural language understanding to user semantic parsing directly limits the performance of the system.Therefore,this thesis focuses on the natural language understanding part of the domain task-oriented dialogue system.Natural language understanding can be split into two sub-tasks: intents recognition and semantic slots filling.Although a large number of research work has been carried out on both tasks,less work has focused on the multiple rounds interaction characteristics of the system.The domain knowledge has not been effectively integrated to improve the performance of the language analysis.In view of the above problems,based on the deep learning model,this thesis studies the intent recognition method combined with context and the slot filling method with domain knowledge.For the multi-round conversation sequential data,an intention recognition model based on context information is proposed.A variety of intentions in a single dialogue are identified by multi-label classification method.The semantic information is modeled at the level of current dialogue and historical dialogue,respectively.The complementary advantages of convolution neural network and bi-directional recurrent neural network are exerted.The experimental results show that compared with the benchmark algorithm,the proposed model has a significant improvement in the accuracy of the intent recognition task.According to the vertical domain-oriented characteristics of the system,a semantic slot tagging model based on domain knowledge is proposed.The embedding layer integrates domain knowledge and user intention.Using the attention mechanisms to learn the contribution of relevant information to the current task in the model.Using the conditional random field to optimize semantic slot sequence tagging results.Compared with the reference sequence marking scheme,the experimental results show that merging domain knowledge and user intention can improve the accuracy of slot identification,and the model has the ability to capture effective information.Based on the research work of this thesis,the proposed model is applied to the movie tickets booking scene.The movie tickets booking task dialogue system based on the domain knowledge graph is designed and built.The movie knowledge graph is constructed as the system data support.The dialogue management is driven by the strategy.The natural language replies are generated by templates.The complete data transfer process within the system has been implemented.