Research On Dialogue Intention Recognition Based On Text Representation And Small Sample Learnin

With the development of artificial intelligence,human-computer interactive dialogue system is becoming more and more popular,in which the intention recognition module has received more and more attention.At present,most intention recognition methods are supervised coding,but in real life,annotation data is often less,and manual annotation cost is high.It is a hot topic in the field of natural language processing to solve the problem of unlabeled small sample data by unsupervised method.Through effective text representation and small sample learning,this study solves the problem of intention recognition in e-commerce customer service dialogues.In this paper,based on the real conversation data of an e-commerce platform,combining text representation and small sample learning methods,comparative learning is introduced to solve the problem of missing positive and negative samples under unsupervised conditions,so as to improve the intention recognition model and apply it into the conversation system of ecommerce platform to improve the accuracy of user intention recognition.Firstly,a two-stage(TF-IDF-MRMR)feature word filtering method is adopted to filter out the helpful feature words to avoid information redundancy among the feature words.The experimental results show that the model recognition accuracy can be improved by 1 percentage point through feature word selection.Then,BERT is used as a text representation for intention recognition,and unsupervised training is performed by contrast learning to make full use of unlabeled data as prior information,so as to achieve text vector representation.By comparing popular encoders such as DPCNN,Text CNN,Text RCNN,Text RNN,Attention and Transformer models,it is found that recall rate,accuracy rate and F1 value increase by 2 percentage points on average.Finally,the metric learning framework is used to construct a small sample intention recognition model.Based on the above text representation as the encoder in the framework,inductive network and relational network are used as the dividers.The overall model is established by optimizing the distance between samples and category representation through parameter learning,and a good intention recognition model is obtained.Compared with other intention recognition models such as Bilst M-ATTENTION and RNN-LSTM,the accuracy of the proposed intention recognition model can be improved by 3 percentage points.Through data processing,model construction and comparative experiments,this paper verifies the effectiveness of text representation and small sample learning for dialogue intention recognition,and expands the application scenarios of small sample learning tasks.