MAIRS: Multi-Agent Collaborative Research Institution Entity Disambiguation with Multi-Source Evidence Aggregation
Year: 2026
Type: conference-paper
Abstract: Research institution entity disambiguation (RIED) is fundamental to scholarly data governance and scientific knowledge graph construction. In real-world scholarly data workflows, institution mentions commonly involve abbreviations, historical aliases, non-standard spellings, hierarchy ambiguity, name collisions, and newly established institutions. We present MAIRS, a collaborative framework comprising Retriever, Linguist, Knowledge, and Context evidence agents and a Judge agent. The four agents produce structured evidence in parallel and may propose missed candidates; a deterministic Candidate Manager within the Judge validates and merges proposals before the Judge aggregates multi-source evidence and triggers debate and reflection when evidence conflicts or is insufficient. We construct 1,200 evaluation instances from real Science Data Bank project records, using manually verified ROR IDs as gold labels. MAIRS achieves 88.4% top-1 accuracy, outperforming string similarity (40.8%), ROR top-1 retrieval (46.5%), and DeepSeek zero-shot inference (77.5%). Candidate Recall@5 improves from 79.3% with initial ROR retrieval to 82.1% after Web expansion and 88.6% after multi-agent candidate proposals. Ablation results confirm that all four evidence agents contribute, with Retriever having the largest effect. Mean latency is 32.14 s per instance, indicating a remaining accuracy-cost trade-off.
Authors: Xiaotong Zhao, Chengzan Li
Language: English
Topic: Data Quality and Management
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7223369652
