One-Class Outlier Detection of Label-Induced Ambiguity in Knowledge Graphs

Knowledge graphs (KGs) such as DBpedia often contain systematic type errors caused not by missing facts but by lexical ambiguity: entities with proper names (e.g., musical bands, natural places, animals) are mis-typed as Person. We term this phenomenon Label-Induced Class Ambiguity (LICA). We propose DeLICA (Detecting Label-Induced Class Ambiguity), an interpretable, pattern-based method that detects such mis-typings by combining predicate with lightweight ontology cues. DELICA learns a compact bank of latent predicate patterns, composes entity representations from predicates and ontology-derived features, and refines labels through a conservative Expectation-Maximization (EM)-style loop with instance-weighting and ranking regularization. It outputs concise pattern-level explanations to support human curation. We formalize the detection problem, connect pattern purity to recoverability under mild assumptions, and introduce practical diagnostics. On DBpedia-style benchmarks, DELICA improves detection quality over lexical heuristics, and unsupervised baselines, while remaining interpretable and practical.