CQL-GNN: Coupled question-label graph neural networks for multi-label educational question classification

Accurate tagging of educational questions with multiple knowledge labels is crucial for personalized learning and resource recommendation. However, multi-label question classification faces significant challenges: labels exhibit long-tail frequency imbalance, many labels overlap semantically, and questions often lack detailed solution context. In this paper, we propose a novel Coupled Question-Label Graph Neural Network (CQL-GNN) framework to address these challenges. CQL-GNN models rich relationships between questions and labels as well as among labels themselves in a unified graph structure, going beyond prior methods that either treated labels independently or only modeled label-label correlations. Each question and label is represented as a node in a heterogeneous graph, enabling dynamic message passing that propagates semantic information from questions to labels and back. This coupled graph approach allows the model to capture label co-occurrence patterns and contextualize label semantics within each question’s content. We integrate pre-trained language models to encode textual features of questions and labels, and design a two-stage propagation mechanism that iteratively refines question and label representations. Experimental results on four real-world education datasets demonstrate that our method consistently outperforms state-of-the-art baselines in terms of Precision@K and F1 scores. Notably, CQL-GNN excels at disambiguating semantically similar labels and significantly enhances the prediction of rare, long-tail labels. The proposed framework provides a robust and generalizable solution for automatic tagging of educational content, with strong potential for adaptation to other multi-label text classification tasks.