Comparative Evaluation of Graph, Hierarchical, and Relational Structuring Paradigms in Large Commonsense Knowledge Bases

Abstract

Commonsense knowledge bases (CSKBs) serve as foundational infrastructures for enabling machines to reason about everyday concepts. While numerous structuring paradigms exist, their comparative efficacy in balancing expressivity, computational efficiency, and scalability remains underexplored. This paper presents a systematic evaluation of graph-based, hierarchical, and relational database approaches for organizing large-scale CSKBs. Graph representations model knowledge as nodes and edges, enabling flexible traversal but incurring overhead in path-sensitive queries. Hierarchical methods, such as taxonomies, optimize inheritance reasoning but struggle with cross-domain ambiguity. Relational paradigms, grounded in formal set theory, enforce strict normalization constraints that enhance integrity but limit dynamic expansion.   We formalize each paradigm using algebraic structures, predicate logic, and complexity-theoretic metrics. A novel hypergraph-based hybrid model is proposed to mitigate rigidity in relational systems while preserving hierarchical inheritance. Experiments on ConceptNet, WordNet, and a proprietary dataset quantify throughput for subsumption, adjacency, and transitivity queries. Results indicate relational systems outperform others in conjunctive queries (F1: 0.92 vs. 0.78 for graphs) but exhibit exponential latency growth during schema revisions. Graph-based approaches achieve high recall in path queries but require heavier indexing. Hierarchical systems optimize memory usage but suffer precision loss in multi-inheritance contexts. The hybrid model reduces transitivity errors by 37% via constraint-driven edge weighting. This work offers framework-agnostic optimization guidelines for CSKB engineers.