MOIE-KG: A multi-agent framework for ontology-guided open information extraction and knowledge graph construction

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Domain knowledge for constructing knowledge graphs (KGs) is often distributed across long documents, where entities and relations span multiple paragraphs and the required type space evolves with document semantics. Constructing KGs from such documents introduces two primary challenges: maintaining global semantic consistency and accommodating document-specific types. Predefined ontology-based methods offer structural stability but lack flexibility, whereas open information extraction techniques enable broader semantic discovery but often yield fragmented structures. Consequently, neither approach adequately resolves the tension between structural consistency and semantic extensibility in long-document contexts. To address this problem, we propose MOIE-KG, a collaborative multi-agent framework comprising specialized agents for coordination and memory management, global information processing, ontology control, entity and relation extraction, and verification and repair. By treating the ontology as a dynamic semantic intermediate representation, MOIE-KG integrates document-level memory, evidence-based feedback, and controlled type projection to preserve structural integrity while facilitating semantic expansion. Experiments on four datasets show consistent F1 improvements over the evaluated baselines. In the fixed-ontology setting, MOIE-KG improves F1 across datasets and backbone models; in the open-ontology setting, it attains a semantic containment rate (SCR) of 86.95%, an entity fidelity (EF) of 94.32%, and a relation fidelity (RF) of 90.37%. These results indicate that MOIE-KG can construct compact, semantically coherent KGs from domain-specific long documents while balancing structural consistency and semantic extensibility.