This paper addresses persistent failures of parsers and relation extractors on long-distance relations (LDRs) by introducing a neurosymbolic Graph Neural Network (GNN) framework that (1) learns multiplicative edge reweighting factors to amplify multi-hop supporting paths, (2) couples a differentiable neurosymbolic constraint loss to enforce linguistic invariants, and (3) provides theoretical and empirical analysis of signal amplification and stability. The method integrates with structured alignment and calibration primitives from the program’s earlier work, improves recall on LDRs while preserving precision, and includes algorithms, proofs (appendix), and a comprehensive experimental plan.