Sentiment Analysis Of Weibo's Comments For Covid-19 Based On Deep Learning

Emotion classification is one of the hottest tasks in the field of natural language processing.Emotion analysis of the comments on public opinion related topics is conducive to controlling the trend of public opinion and maintaining social stability.In the background of new epidemic COVID-19,100 thousand comments data on Sina micro-blog platform were collected from January 1,2020 to February 20,2020 to classify emotions.During the data preprocessing stage,data were manually annotated and data cleaning methods were used to denoise.In the part of model building,an emotion classification model with good classification effect is built based on multiple deep neural network structures and transfer learning.The model uses the dynamic pre-trained model Bidirectional Encoder Representations from Transformers(BERT)for preliminary feature extraction and ‘word vector' coding method.On these basis,the two-way Long-Short Term Memory(LSTM)model and attention mechanism model are used to extract the context information and important information of the text.After the output vector is spliced with the output vector of the classification(cls)position of the BERT model,it is input into the full connection layer for emotion classification.The experimental results show that,compared with the benchmark models BERT splicing Convolutional Neural Networks for Text(Text CNN)and BERT splicing Deep Pyramid Convolutional Neural Networks(DPCNN),the model built in this paper has achieved an accuracy of 85.54% and 0.8543 for Macro-F1 in emotion classification on the micro-blog comment data at the initial stage of COVID-19.The results show that the proposed model has better emotional feature extraction ability on micro-blog comment data for COVID-19,and it is suitable for three classification emotional tasks.