A Novel Graph-Based Approach Using Graph Neural Networks for Structuring Large-Scale Commonsense Knowledge Bases

Abstract

Commonsense knowledge bases (CSKBs) represent real-world facts through entities and relationships, yet their often unstructured or semi-structured nature impedes efficient scalability and reasoning. This paper proposes a novel graph neural network (GNN)-based framework designed to autonomously organize large-scale CSKBs into semantically coherent, machine-interpretable graphs. The core architecture incorporates heterogeneous graph convolution operators to simultaneously capture multi-relational dependencies and preserve local topological features. In tandem, a hierar- chical attention mechanism adjusts edge weights dynamically based on node attribute similarity and global graph connectivity, leading to efficient discovery of sparse substructures. To mitigate the combinatorial explosion of relational paths, we introduce a constrained optimization objective that minimizes edge reconstruction loss while maximizing deductive closure through implicit Horn clause satisfiability. Empirical evaluations on ConceptNet and ATOMIC confirm notable improvements in multiple metrics: a 23.7% boost in edge prediction accuracy over competitive graph autoencoder baselines and a semantic consistency score of 0.892 on held-out triples, outperforming transformer- based knowledge base completion methods by 15.2%. Qualitative structural analyses reveal emergent hypernymy and causality hierarchies without explicit ontological supervision, highlighting the robustness of the learned graph representations. Our results emphasize that unifying geometric embeddings with symbolic reasoning constraints significantly enhances the structuring of noisy commonsense assertions, thus promoting scalable, high-fidelity CSKBs.