OntoDup: Governance-Aware Entity Matching for Scholarly Knowledge Graph Deduplication

Scholarly knowledge graphs integrate bibliographic records from heterogeneous sources and therefore require controlled, auditable deduplication. This paper presents OntoDup, an ontology-driven approach that models entity matching as a governed decision process: Matching outcomes are recorded as reified assertions enriched with governance state, evidence, provenance and operational metadata, while a separate operational view is exposed through policy-driven materialization of consumable identity links. We evaluate OntoDup on the DBLP-ACM and DBLP-Scholar benchmarks under two regimes: (i) a pre-blocked setting using the benchmark candidate lists to compare matching methods under a fixed candidate set, and (ii) an end-to-end setting that generates candidates from the graph with DeepBlocker and applies governed triage and materialization. We report operational precision/recall/ (Formula presented.) computed directly on the graph via SPARQL aggregations, characterize governance workload through state distributions, and quantify inference cost for LLM-based matchers via token and latency metadata attached to assertions. For end-to-end evaluation, we anchor operational links against a full positive reference encoded as idealized validations derived from the benchmark labels, enabling analysis of missed positives in terms of governance status and materialization policy. The experiments show that OntoDup enables evaluation at the level of consumable identity links, review workload, and inference cost, revealing operational trade-offs that are not visible from pairwise matching metrics alone.