Research On Intent Recognition Based On Task-based Multi-round Dialogue

With the rise of artificial intelligence and the development of mobile internet,the dialogue system,as a new type of human-computer interaction,has gradually come into people's lives and received continuous attention from academia and industry.Intention recognition is a key subtask of natural language understanding in human-computer dialogue system.It is mainly used for the dialogue system to understand user's language correctly.Its accuracy affects the reasonable reply generation of the dialogue system and the service quality of the entire dialogue system.In recent year,deep learning models has a better performance in single intention recognition tasks in recent year.However,how to use the historical information of current dialogue to understand its multiple intentions is the difficulty of dialogue system.In view of the problem of multi-intention recognition in multiple round dialogue,the specific work is as follows:First,for the existing intent recognition method,its word vector representation of the text cannot solve the problems of synonyms,polysemy,etc.Thus,Bert pre-training model is selected in this paper to obtain character-level vector representation of text,getting richer semantic information.Second,the existing intent recognition model has a single use of historical information.In this paper,a classification model based on bidirectional gated recurrent unit and attention mechanism is designed for a variety of typical dialogue scenarios.In this model,four kinds of historical information introduction methods are proposed and control unit are added to screen,which further improves the classification effect.Third,for questions where the dialogue text may contain multiple intents Thus,in this paper,a multi-intention recognition model based on an improved capsule network is designed.Based on the historical information and control unit,the model replaces the convolutional layer of capsule network by a bidirectional gated loop unit,and adds a self-attention layer and a convolution capsule layer to extract key semantic information of the intention text,improving the performance of multi-intention classification.Experiments,which is on the datasets of single intention and mixed intention in multiple dialogues,show that the micro-average F1 values increased by 12.91% and 6.79% respectively after the introduction of historical information and control unit.The microaverage F1 value of the improved capsule network model increased by 14.60% and 13.67% respectively.