12 / 2026-10-30 23:50:29

PorosKG: Structure-Guided Knowledge Graph Construction over Scholarly Literature

Scholarly knowledge graph,Entity resolution,Cross-document entity consolidation,Citation graph

Constructing scientific knowledge graphs (KGs) from large scholarly corpora requires consolidating entities across interlinked documents. Terminology is context dependent, and exhaustive semantic matching is prohibitively costly when each comparison relies on rich descriptions or LLM reasoning. Existing approaches mainly rely on semantic similarity, while citation links are typically treated as undifferentiated connections. We propose PorosKG, a structure-constrained framework that uses the explicit organization of scholarly literature to narrow the candidate space for cross-document entity consolidation. PorosKG models intra-document hierarchy to preserve local entity context, and refines citation links into three consolidation-oriented relations: inheritance, parallel, and peripheral. Description-based matching is then performed only for entity pairs linked by inheritance or parallel relations, with the preserved context used for disambiguation. Experiments on public scientific benchmarks and an expert-annotated NanoScience corpus show that PorosKG achieves a CoNLL F1 of 0.66 in cross-document coreference evaluation, outperforming the strongest evaluated baseline by approximately 5.0 points. In the ablation study, PorosKG reduces token usage by approximately 82.2% relative to Global Semantic Match, with an entity-level F1 of 0.82 compared with 0.89 for global matching. Further ablations show that typed citation relations improve entity-level F1 from 0.75 to 0.82 compared with untyped citation links.