Sentiment classification optimization based on improved Senti-BERT and BiLSTM-Attention

This study focuses on the problem of insufficient accuracy in sentiment classification in complex texts and multimodal scenarios. A sentiment classification optimization method is proposed, which integrates a bidirectional encoder representation Transformer model for sentiment analysis (Senti-BERT)and a bidirectional Long Short-Term Memory with Attention (BiLSTM-Attention). Firstly, the Senti-BERT layer is introduced to optimize the generation of word vectors by introducing an emotion dictionary leading mechanism, and a dynamic weighting strategy is adopted to fuse multi-modal features such as text, emoticons, and images. Subsequently, the hierarchical attention mechanism of BiLSTM-Attention structure is used to enhance contextual modeling. Finally, the stacking method is adopted to integrate the advantages of multiple models. The experimental results show that this method performs well in multiple dimensions. This optimization method achieved an F1 score of 84.6% in verb sentiment word recognition and 90.1% in adjective recognition, which is 4.8% higher than the bidirectional encoder representation Transformer. The full modal fusion results in a classification accuracy of 89.4%, and the English accuracy exceeds 90% in cross language tasks. The accuracy of emotion segment localization using dynamic weight attention mechanism reaches 90%. The quantization model maintains a low latency of 92ms under 10000QPS concurrency, and the Stacking strategy increases the AUC value to 0.943. The research provides a high-precision solution for social media sentiment analysis through innovative multi-level architecture design.