Research On Social Network Texts Emotion Recognition Based On BERT Model Feature Construction

With the rapid development of the Internet and the rise of new media forms of social platforms,countless users participate in the use of social platforms.As the text information with strong emotional color,the text of social network is of great help to the research of users' interest orientation.How to accurately identify and analyze the emotional color of these text information has become the main research problem of many researchers in the current academic field.In recent years,the BERT model,which has been trained for a long time in an unsupervised way,has achieved remarkable results in the field of NLP.Based on the BERT model,this paper proposes a feature construction method called B-feature,which combines the improved LSTM model to analyze the text information of social networks.The main contents of this paper are as follows:(1)Segmentation,as the main step of traditional methods to obtain text vector,has some limitations.In this paper,we propose a method of feature construction based on the BERT model,B-feature,which preserves the main feature information of text sentence on the basis of avoiding word segmentation.After the information of sentence coding is supplemented,recombined and deleted,the feature vector of the text is constructed on the basis of the obtained sentence coding,and the feature constructed based on the BERT model after processing is obtained.Compared with TF-IDF and other methods,the accuracy of the proposed algorithm is improved by 7.8%,3.8% and 11.7% respectively in the case of the same neural network classifier.(2)Aiming at the problem of low accuracy of traditional methods for text emotion recognition in social networks,this paper selects LSTM as the experimental basic model.On the basis of one-way LSTM,a reverse LSTM is added,and the attention mechanism is introduced to make it pay more attention to the key features in the features.Then,the feature construction method B-feature and the two-way LSTM which introduces attention mechanism are combined to build the model,and the feature information after the feature construction is sent to the attention mechanism LSTM for training and experimental classification.Experiments show that the accuracy of the proposed model is 2.96% and 3.87% higher than that of the traditional method combined with LSTM.(3)Based on the algorithm and model combination proposed in this paper,the hot topics analysis system of social network is designed and implemented.This system obtains hot topics and topic comments on social platform through crawler,and carries out emotional recognition on the acquired text information through the model of this paper,so that the emotional color of social network text can be presented to users with visual interface intuitively.It not only has comprehensive functions,but also improves the efficiency of information acquisition and realizes the application values of this paper.