Research On Multi-Emotion Classification Based On Att-Blstm-Mc Model

Most emotion classifications in the past were based on coarse-grained granularity,and there was not much research on the five classifications.The final emotion classification results only contain one emotion and few studies on the coexistence of multiple emotions,so it cannot fully reflect the richness of the user's emotions.Research scholars have also achieved good results in the study of monolingual or bilingual texts using the Attention mechanism,but few studies have focused on the emotion classification of Code-Switching Text.To address these shortcomings,this paper proposes a multi-emotion classification method based on Attention and BiLSTM model.The BiLSTM model is used to construct five emotion classifiers to obtain text context relationships,namely Happiness?Anger?Sadness?Fear and Surprise emotion classifier.To predict all emotions that a single post belongs to.Use word2vec's Skip-gram method to convert Code-Switching Text into word embedding as input to multi-emotion classifiers,and use negative sampling method to improve the quality of word embedding and speed up training,introducing Attention for different weights of different words on text Mechanism to express the importance of different features,enhancecontextual semantic information,and obtain deeper features,and finally return through Softmax to complete all emotional predictions of all posts.By analyzing the experimental results based on Attention and BiLSTM model,it is found that the performance of Surprise and Fear emotion classifier are significantly lower than that of Happiness?Anger and Sadness emotion classifiers.The reason for the analysis is that the Code-Switching Text has category imbalance problems.Because the categories with small data features are not obvious when training the model,the categories with small data are predicted as the categories with large data,which reduces the accuracy of emotion classification.In order to solve the above problems,this paper proposes an improved multi-emotion classification method based on Attention and BiLSTM models.Use the API provided by Sina to capture the Code-Switching Text data in Weibo,invited 20 students in the field of natural language processing to mark the data category of the captured data,and select five types of data from the target,such as selecting the same Posts contain a variety of emotion data,the amount of data in each category is unified to 2000,so that the Code-Switching Text reaches the category balance and expand the Code-Switching Text.Pseudo gradient descent method is used to adjust the model parameters,optimize the cross-entropy loss function,reduce the complexity,and make the classification performance of the model more stable.The experiment proved that the improved multi-emotion classification method based on Attention and BiLSTM model compared with the unimproved algorithm,alleviates the problem that the minority class is misclassified into the majority class,the Marco-F1 value is increased by 11.4%,and the F1 value of the each emotion classification classifier has been greatly improved,reducing the gap between the F1 value of the Surprise and Fear emotion classifiers and the F1 value of the other three emotion classifiers.Therefore,the improved multi-emotion classification method based on Attention and BiLSTM model can more accurately predict multiple emotions of the text,achieve better classification effect,and verify the effectiveness of the model.