Research On Multi-classification Method Of Micro-blog Negative Emotion Based On Deep Learning And Meta-learning

With the continuous development of network technology and social media platforms,Weibo has become the largest media communication platform in China.Netizens usually communicate with each other in written form,offering their opinions on events,goods and services.While delivering information efficiently and conveniently,these messages often contain a large number of users’ emotions.Therefore,if these emotional text data can be analyzed,it can provide a lot of valuable information reference for the study of user preferences,understanding of social value tendency,improving the quality of goods and services and other aspects,which has great social significance and commercial value.However,the study of emotion problems cannot be realized only by artificial methods,so it promotes the continuous development of natural language processing technology,making emotion analysis technology become a research hotspot.In the face of new tasks,traditional sentiment analysis algorithms often need a large amount of data to be retrained in order to get better accuracy.In view of this situation,this thesis proposes a multi-classification method of micro-blog negative emotions based on MAML and BILSTM.Firstly,according to the experimental requirements,the negative emotion data set of Weibo was created,and the data were preprocessed to a certain extent to construct a word segmentation dictionary suitable for the experiment in this paper.Then,a combined model of MAML and BILSTM was constructed,in which BILSTM realized sentiment classification of micro-blog data.MAML updates the meta-learning parameters through gradient descent by calculating the sum of losses from multiple training sessions.Experimental verification shows that compared with the current popular model,the accuracy rate,recall rate and F1 value in the negative emotion data set of Weibo are increased by 1.68%,2.86% and 2.27% respectively.At the same time,in order to further enhance the model in weibo more negative emotion classification task performance,this paper proposes a word vector representation method based on improved BERT,combined with statistical knowledge and BERT model,using the optimized term vectors as input of classification model,experimental results show that the optimized word vectors to weibo more negative emotion classification with lifting effect,on different models,the accuracy and recall rate,F1 highest value increased by 1.49%,1.25% and 1.04%,respectively.