Text Classification for Social Governance: A Novel Strategy with Adaptive Reward Mechanisms

Given the increasing need for societal governance systems to effectively address public grievances, classifying texts from public service hotlines and other governance communications has become challenging, especially for texts with complex hierarchical structures. This study explores how adaptive reward mechanisms can be applied to classify complaint texts in societal governance. We propose a novel hierarchical text classification model that overcomes the limitations of traditional methods in handling hierarchical data. We designed a hierarchical loss function that considers inter-label hierarchical relationships to enhance the model's learning of hierarchical information. Additionally, we used a simple label encoding method to reduce computational complexity and improve training efficiency. Experimental results show that our model significantly outperforms existing methods in classification accuracy, both flat and hierarchical. Furthermore, ablation studies confirmed the hierarchical loss function's crucial role in enhancing model performance. This research offers an effective strategy for classifying complex hierarchical data, significantly improving classification tasks in natural language processing.