Bullet screen comments on interactive video platforms are a rich source of real-time emotional signals. However, their anonymity, brevity, and contextual dependence pose significant challenges for analysis. To address the challenges, we developed a virtual sensing architecture tailored to bullet screen data, integrating abnormal user detection, fine-grained emotion inference, and thematic clustering. A random-forest-based denoising module achieved 91.99% accuracy, effectively filtering malicious or noisy accounts and securing high-fidelity input streams. For emotion recognition, the improved recurrent convolutional neural network for text (TextRCNN) model incorporating pretrained word vectors and enhanced feature fusion showed a 71.53% accuracy, outperforming baseline models such as support vector machine (58.2%), TextCNN (65.4%), and standard TextRCNN (68.1%). The model demonstrated strong recognition of distinct emotions, particularly joy and anger. To address short-text sparsity, the biterm topic model was constructed, yielding a coherence score of 0.58, significantly higher than latent Dirichlet allocation at 0.41, and successfully clustering comments into interpretable themes such as character discussion, production evaluation, and criticism. Results were visualized on a real-time monitoring dashboard, enabling platform-wide emotional health assessment. By translating noisy social signals into high-fidelity emotional and thematic information, the developed architecture supports the digital-twin construction of collective sentiment and advances sensor technology for next-generation interactive media systems.