A BERT-based Text Sentiment Classification Algorithm through Web Data

In order to analyze the sentiment tendency of public opinion, this paper conducts a textual sentiment classification research through web data. In the research, this paper uses the BERT (Bidirectional Encoder Representation from Transformers) model to replace the commonly used word2vec model as a text vectorization tool, which has stronger semantic representation capabilities and can realize polysemous words. For the multi-label classification problem of reviews, the BR (Binary Relevance) algorithm is used to transform the problem into multiple binary classification problems, which is directly and efficient for processing multi-label data. Design the BiLSTM-Attention model, which combines the bidirectional long and short-term memory network and the attention mechanism to achieve further extraction of text features. After multiple sets of comparative experiments, the effectiveness of the BiLSTM-Attention model is verified through performance evaluation. In order to further improve the performance of the model, the problem of unbalanced data set is solved by adjusting the loss function and various parameters so that a better classification effect is achieved.