Abstract
Sentiment classification on Weibo has recently attracted wide attention in the research community. Most previous works focus on comments regarding movies or products. This study, in contrast, targets "gusty incidents" on Weibo—sudden, high-profile events that trigger surges in public opinion. We propose an innovative convolutional neural network-based method, termed CNN-SVM, to classify incident comments. Additionally, we introduce a new data structure, the "repost tree," to handle ambiguity in comments by tracking user repost actions. Extensive experiments demonstrate that the CNN-SVM method effectively improves classification accuracy, while the repost tree aligns results with real-world sentiment tendencies.
Introduction
Microblogging platforms like Sina Weibo are hubs for public opinion during sudden incidents. Analyzing sentiment during these events is challenging because posts are short, informal, and often lack context. A significant hurdle is the repost mechanism: a user might "agree" with a negative post, making their literal text appear positive while their underlying sentiment toward the incident is negative. This paper addresses these challenges through a hybrid machine learning model and a structural approach to data.
The CNN-SVM Hybrid Model
The proposed CNN-SVM model combines the feature extraction strengths of Convolutional Neural Networks (CNN) with the robust classification capabilities of Support Vector Machines (SVM).
As shown in the architecture, the CNN layer uses word2vec embeddings to extract local features from short texts. Instead of relying solely on a standard softmax layer for final classification, the high-dimensional features extracted by the CNN are fed into an SVM classifier. SVMs are generally more robust for high-dimensional, small-sample pattern recognition, providing a performance boost over traditional CNN architectures.
The Repost Tree Mechanism
To capture the contextual logic of shared information, the authors introduced the Repost Tree. This hierarchical structure tracks the path of a repost. Each node is assigned a sentiment value (+1 for positive, -1 for negative). The "real" sentiment of a specific repost (a leaf node) is calculated as the logical product of all nodes in its branch. For example, if a user "agrees" (positive action) with a "negative" post, the logical product correctly identifies their stance as negative.
Experimental Results and Real-World Alignment
The model was evaluated on three datasets: MN (standard), WZ (large/noisy), and a specific dataset regarding the 2015 Paris attacks. The CNN-SVM model consistently outperformed standard SVMs, Deep Belief Networks (DBN), and standard CNNs, achieving 93.28% accuracy on the MN dataset.
The effectiveness of the Repost Tree is most evident when comparing model predictions to real-world data. In the Paris incident case study, text-only models deviated significantly from public sentiment ratios reported by People's Daily. By incorporating the Repost Tree, the model "steered" the results to within 7.46% of the actual public sentiment ratio, as illustrated in the comparison above.
Conclusion
By combining CNN-based feature extraction with SVM classification and introducing the Repost Tree for contextual logic, this research provides a more accurate and representative method for sentiment analysis on social media. Future work will focus on addressing linguistic nuances like sarcasm, which remain a challenge for logical product calculations.
Summary updated