Research And Application Of Intention Recognition And Response Generation Technology Based On Sequence

With the development of natural language processing technology,the application of taskoriented dialogue systems become more and more extensive.Intent recognition and response generation are two important components of task-oriented dialogue.Intent recognition is designed to infer the target of the user's current action by understanding the user's context.Response generation generates dialogue content feedback to the user based on the current dialogue state which makes the dialogue content more consistent with the current scene of the dialogue.Therefore,the research on intent recognition and response generation is helpful to build dialogue systems more similar to what user needs.Currently,intent recognition and dialogue generation still face some challenges.On one hand,the multiple rounds of dialogue history between the system and the user cannot be fully applied in the intention recognition model.At the same time,the user's intention is difficult to infer since the user's intention would change along with the dialogue processing.On the other hand,the generated reply content often conflicts or overlaps with the dialogue history due to the insufficient perception of historical information by the generative model.Started from the issues above,we carry out the following work:(1)Aiming at the characteristics of intention sequence correlation and situation of mixed multi-modal information existing in intelligent system dialogues,we propose a kind of Multimodal Dialog Intention Recognition with Sequence Modeling.The model combines the image information existing in the dialogue process,captures the user's intention changes through the dialogue sequence,and introduces the graph attention network to optimize the coding of the intention sequence.Experimental results in a real E-commerce scenario shows that the introduction of the intent sequence modeling with the graph attention network and multi-modal information can make the model more accurately identify the user's current dialogue intention.(2)Aiming at the problem of existing response generation models ignore the important role of a historical dialogue action sequence in dialogue history perception.We propose a kind of Dialogue Action Sequence Based Co-Generation Model.In this model,dialogue action prediction and response generation are regarded as multi-task learning,to which historical dialogue action sequence information is introduced.The experimental data on the multi-round dialogue data Multi WOZ show that the generative model after introducing the historical dialogue action sequence has superiority in the comprehensive performance of various indicators.(3)This paper applies the idea of sequence modeling and the response generation model in the above work to psychology and develop an emotional support dialogue system based on dialogue strategies.The system customizes the dialogue strategies under the guidance of Helping Skills theory and selects the appropriate strategy to guide the response generation through the dialogue history.The system framework is implemented based on Django,and the Pepper robot is used as a typical terminal for deployment,which can provide users with effective emotional support when their emotions are negative.