Dual-Layer Multi-Head Attention Enhanced RoBERTa-BiLSTM for Text Sentiment Classification

With the exponential growth of User-Generated Content(UGC) on e-commerce platforms, sentiment analysis of product reviews has become a pivotal technology for enhancing the intelligence of business decision-making. However, existing models face challenges such as semantic ambiguity, dispersed sentiment cues, and domain-specific vocabulary when processing Chinese product reviews. To address these issues, this paper proposes a Dual-Multi-Head-Attention-RoBERTa-BiLSTM model(DMHA-RoBERTa-BiLSTM). Firstly, the model generates dynamic context-aware word embeddings using the RoBERTa-wwm-ext-large pre-trained model. Secondly, a dual-layer multi-head attention mechanism is designed to extract features through global contextual correlations and local emotional keyword focusing, synergized with a BiLSTM network to model long-distance affective dependencies. Finally, it achieves sentiment polarity classification through feature fusion. Experiments on a Chinese e-commerce review dataset demonstrate that our framework attains 90.8% classification accuracy with comparable F1 performance (90.3%), outperforming baseline models such as BiLSTM-CNN and HF -MHA by 2.1%-7.5%. Ablation studies further validate the critical contribution of the dual-layer attention architecture to the model's performance.