Sentiment Classification of Social Network Text Based on AT-BiLSTM Model in a Big Data Environment

Abstract

To tackle the challenge of ineffective sentiment prediction using current sentiment classification methods, this paper introduces a method social network text sentiment classification. The method leverages a bidirectional short and long-term memory model (AT-BiLSTM), specifically designed for a big data environment. First, a vectorized representation of text is realized by introducing a pre-trained BERT model, and the classification results are dynamically adjusted according to the semantic information of the words. Then, the BiLSTM combined with the attention mechanism performs aspect-level sentiment analysis, and the corresponding model AT-BiLSTM is formulated. Finally, the BERT model randomly selects input tags for information masking and pre-trains the proposed model. The proposed method was evaluated against three alternative methods using an identical dataset. The results show that the novel method achieved the highest accuracy, recall, and F1-score, reaching 93.72%, 93.91%, and 92.38%, respectively. Consequently, the proposed method demonstrates superior performance compared to the other three methods evaluated.