Research on Sentiment Analysis of Weibo Comments Based on ALBERT-BiGRN-Att
Aixiang He, Xiqin Ao, Weiwei Wu, Wu Liru · 2025
The widespread use of the Internet since its hasty development has brought social media platforms such as Weibo, where the public reads and exchanges their opinions and expresses their emotions. To solve the complex and implicit sentiment analysis problems of short-text Weibo comments, this paper proposes a deep learning model called ALBERT-BiGRN-Att, which incorporates a pre-trained language model and a hierarchical attention mechanism. In this model, a tri-level synergistic architecture is incorporated for better feature extraction: (1) dynamic semantic encoding to learn context-sensitive word embeddings through ALBERT; (2) bidirectional temporal dependency modeling through a bidirectional gated recurrent network (BiGRN) to enhance contextual representation; (3) hierarchical attention mechanisms acting on crucial emotional keywords at the word level and key semantic units at the sentence level which indeed augment the model's ability to capture subtle emotional trends to a great extent. We conduct experiments in a hybrid set, which consists of the SMP2020 Weibo Sentiment Classification benchmark and actual trending comments crawled by Scrapy. Five-fold cross-validation results show that the proposed model achieves state-of-the-art performance accuracy of 84.7%(six-class including happiness, anger, sadness, sadness, surprise, neutral) and 84.3 %(three-class including positive, negative, neutral) and outperforms various baseline models(GRU, BiGRU, BERT, etc.) with average margins 5.3 and 3.56 percent with macro-F1 as 85.6% and 84.8% respectively. Ablation studies also demonstrated the effectiveness of the hierarchical attention mechanism and label smoothing cross-entropy loss, while visualization techniques—e.g., confusion matrices and t-SNE plots—demonstrated robustness of model clustering and decision logic in semantic space. It is hoped that this work would offer a feasible method in dealing with short-text sentiment analysis towards social media, and it has been widely used in application scenarios including public opinion monitoring, user profiling and marketing analysis.