RLHNN: a reinforcement learning-enhanced hybrid neural network for public opinion short text classification

The widespread use of social media has led to the rapid generation of opinion-rich short texts, posing increasingly complex and high-variance challenges for public discourse analysis. These texts are typically informal, context-limited, and linguistically diverse, which creates significant difficulties for traditional static classification models. In this context, short text classification has emerged as a fundamental yet challenging task that lies at the core of understanding and managing opinion-oriented content. To address these challenges, we propose RLHNN, a Reinforcement Learning-Enhanced Hybrid Neural Network that trains a classification head (policy) using an Actor–Critic setup: the Actor outputs class probabilities from fused features, and the Critic predicts the expected reward; both are optimized jointly via a policy gradient on a binary reward. The proposed architecture integrates a multi-scale Transformer for semantic feature extraction and a capsule network for modeling spatial dependencies. A self-attention layer after feature fusion re-weights the concatenated representations. RLHNN achieves 89.74% accuracy on the Toutiao News dataset and 92.45% on the AG’s News dataset. On our self-constructed social media corpus of complex, ambiguous, and imbalanced short texts, the model reaches 89.00% accuracy. The approach enhances short text classification by effectively combining multi-scale and capsule features with reward-guided policy updates.