Integrated sentiment analysis of social media based on deep learning and topic modeling—a case study of Xiaomi's Weibo trending data

Abstract

As social media serves as a public arena for information aggregation, the volume of user-generated comment data provides resources for corporate public opinion monitoring and product feedback. This paper proposes a sentiment analysis framework that integrates the BERT language model, jieba word segmentation and keyword extraction, and LDA topic modeling, and validates the approach using real-world Weibo comment data. Python web scraping was employed to collect user comments under Weibo trending topics related to Xiaomi from January 1 to March 20, 2025, yielding 12,396 valid samples after data cleaning. A random subset of 2,900 comments was manually annotated for fine-tuning the BERT model to perform three-class sentiment classification (positive, negative, neutral). Jieba segmentation and stop-word filtering were applied, TF-IDF statistics were computed, and LDA modeling was utilized to extract latent topics within positive and negative sentiment categories. The results indicate that the accuracy on the test set reaches 78.85%, confirming that the model exhibits stable classification performance and strong generalizability. BERT-based sentiment analysis reveals that positive sentiments account for the largest proportion. Keyword analysis and topic modeling indicate that positive comments center on "technological innovation" and "personal image"; negative comments concentrate on "product performance" and "brand service"; while neutral comments are factual descriptions of product specifications and event information. The study confirms that the proposed integrated approach effectively addresses the semantic complexity inherent in social media text sentiment analysis, offering an actionable solution for enterprise public opinion analysis in related domains.