Magnetic Polarity: An Automatic Classification Framework for Compatible and Contradictory Relations in Knowledge Graphs

This study proposes the Magnetic Polarity framework, which automatically classifies inter-knowledge relations in knowledge graphs into five polarities: attract (compatible), repel (contradictory), contain (subsumptive), exclude (exclusive), and neutral (unrelated). The framework adopts a 3-Level Cascade architecture (L1 embedding similarity, L2 rule-based, L3 LLM reasoning) that optimizes the cost-accuracy tradeoff. In experiments on 1,200 ConceptNet pairs, L2 rules alone achieved 100% accuracy when relation types were given, while the Cascade maintained comparable accuracy to All-LLM (50.2% vs 51.2%) with 16% fewer LLM calls in text-only conditions. Downstream analysis demonstrated that contradictory knowledge reduced LLM generation accuracy by 8.2 percentage points and contaminated 58.4% of responses.