Event linking aims to associate event mentions in text with entries in a knowledge base or label them as out‑of‑KB events. Existing approaches differ in architecture, yet candidate acquisition remains vulnerable to short ambiguous mentions, noisy arguments, and unevenly useful evidence. To address this, we introduce MACE (Multi‑Agent Candidate Event), which refines the event structure before linking. MACE employs four evidence‑specialized LLM agents to gather time, location, participant, and event‑type evidence; these are transformed into intermediate queries for candidate‑event lookup tools. A coordinator then reviews and revises the evidence set before constructing the final candidates. Experiments on two public event‑linking benchmarks, combined with various baseline models, show that adding MACE consistently improves accuracy, demonstrating that better candidate acquisition alone can markedly boost overall linking performance without altering the underlying linking model.
Review