Task-Aware Text Graph Structure Learning for Semi-Supervised Text Classification

Current text graph structure learning methods often rely on statically constructed graphs, creating a semantic mismatch with downstream tasks and limiting representation quality under computational constraints. To address this, we propose TATGSL, a novel framework that reformulates semi-supervised text classification as a joint optimization of node representations and graph topology. The core of TATGSL is a unified, alternating optimization objective that coherently combines classification, task-aware metric learning, and homophily-consistency loss. This principled integration is stabilized by a confidence-aware co-training schedule, which dynamically refines the graph to align with task objectives while effectively preventing error propagation. Extensive experiments on five benchmarks demonstrate that TATGSL achieves state-of-the-art performance, while maintaining modest memory usage and exhibiting strong robustness. Our code and dataset are available at: https://github.com/WithMeteor/TATGSL .