Lightweight reinforcement graph learning framework for low resource sensitive text classification

Classifying sensitive texts under extreme low-resource constraints is highly challenging due to fine-grained categories and confusable boundaries. While graph neural networks capture structural semantics, they suffer from static topological noise. Furthermore, existing reinforcement learning (RL) enhancements relying on pseudo-labels often trigger severe confirmation bias. To address these limitations, we propose LRG-Sens, a Lightweight Reinforcement Graph learning framework for Sensitive text classification. We formulate graph construction as a Contextual Bandit problem, where an RL agent dynamically prunes vocabulary nodes to physically purge topological noise, thereby inducing an optimal, compact document graph. Bypassing fragile limited labels, the agent is driven by a novel physics-inspired unsupervised reward that harmonizes information energy retention, topological sparsity, and an energy-gated isolation penalty to preserve critical semantic hubs. Subsequently, a hierarchical transductive graph convolutional network leverages this purified topology to simultaneously evaluate labeled and unlabeled nodes, maximizing the utility of sparse supervisory signals. Extensive experiments on our introduced real-world confidential corpus (CIA) and public benchmarks demonstrate that LRG-Sens achieves state-of-the-art accuracy and stability in extreme 5-shot scenarios. It significantly reduces computational overhead while explicitly outperforming massive instruction-tuned 8B large language models on highly obfuscated confidential texts.