Confidence-Aware Dual-Graph Calibration for Noise-Robust Relation Extraction in Low-Resource Scenarios

Relation Extraction (RE) models often struggle in real-world settings with data scarcity and noisy supervision, where limited training samples amplify overfitting to spurious correlations and unreliable evidence. Graph-based approaches can exploit structural cues, yet they face a key trade-off: syntactic dependency graphs provide stable but potentially incomplete (and parser-dependent) structures, whereas learnable semantic affinity graphs are more flexible but prone to noisy connections, especially in low-resource regimes.To address these challenges, we propose Confidence-Aware Dual-Graph Calibration (CDGC), a unified framework for noise-robust RE. CDGC constructs two complementary views over the same sentence: a fixed syntactic dependency graph for structural guidance and a learnable semantic affinity graph for contextual interactions. We introduce a Structure-Guided Semantic Calibration module that injects syntactic priors to regularize dense semantic affinities and suppress spurious semantic edges. Furthermore, we devise a confidence-aware denoising strategy that measures structural--semantic predictive inconsistency via Jensen--Shannon divergence, and uses it to reweight unreliable instances and perform confidence-adaptive nucleus pruning on the semantic graph, without requiring additional clean supervision for denoising.Extensive experiments on TACRED, FewRel, and SemEval-2010 Task 8 under diverse low-resource and noisy-label protocols show that CDGC consistently outperforms strong structural and noise-robust baselines; for example, on TACRED with 10% training data and 30% label noise, CDGC improves micro-F1 from 48.7 to 53.2 (+4.5) over RoBERTa. Beyond extraction performance, we explicitly evaluate probabilistic calibration and show that CDGC yields substantially more reliable confidence estimates, reducing ECE from 0.142 to 0.051 and Brier score from 0.354 to 0.185 on the same setting. Additionally, we evaluate on DocRED to demonstrate that the dual-graph framework generalizes to document-level reasoning under low-resource clean supervision.