Enhancing Event Candidate Acquisition for Event Linking

Ziyang Zhang; Yinan Liu; Boyi Xue; Yingxuan Huang; Bin Wang; Xiaochun Yang

cs.AISep 12, 2026

Enhancing Event Candidate Acquisition for Event Linking

Authors: Ziyang Zhang, Yinan Liu, Boyi Xue, Yingxuan Huang, Bin Wang, Xiaochun Yang

Organizations: School of Computer Science and Engineering, Northeastern University Shenyang 110819, China

Abstract

Event linking associates event mentions in text with entries in a knowledge base (KB), or identifies them as out-of-KB events. Although existing methods use different architectures, candidate event acquisition can still be weakened by short ambiguous mentions, noisy arguments, and evidence that is unevenly useful for retrieval. We present MACE, a Multi-Agent Candidate Event acquisition method that refines event structure before linking. MACE uses evidence-specialized LLM agents to acquire time, location, participant, and event-type evidence, exposes intermediate queries to candidate-event lookup tools, and lets a coordinator revise the evidence set before final candidate construction. Experiments on two event linking benchmarks show that adding MACE to different event linking models consistently improves accuracy. These results show that MACE improves event linking through better candidate event acquisition without modifying the event linking model.

Figures & tables

Figure 1: Overview of our framework MACE.

Method	Wikipedia Test			NYT Test
Method	All	Verb	Noun	All	Verb	Noun
GENRE (ICLR 2021)	26.55	13.48	39.61	23.54	22.94	24.71
GENRE+MACE	43.65	30.09	57.20	42.78	41.76	44.79
EveLink (EACL 2023)	79.22	78.07	79.93	32.03	34.34	25.13
EveLink+MACE	84.80	86.68	82.93	67.29	71.24	55.53
SEFEL (AAAI 2026)	83.64	85.99	81.29	76.90	79.61	68.85

Table 1: Performance on the task of event linking. The reported metric is accuracy (ACC, %). The EveLink Wikipedia value in the K-expansion control is a separately reproduced run; see the accompanying note and Appendix A .

Ablation	Wikipedia Test	NYT Test
Ablation	All	All
SEFEL + MACE	84.83	79.44
w/o Agent Extraction	82.22	56.53
w/o Verification	84.11	76.92
w/o Coordinator	83.76	75.97

Table 2: Ablation study results. The w/o Agent Extraction, w/o Verification, and w/o Coordinator rows remove the corresponding extraction, verification, and coordinator-loop components while keeping the downstream SEFEL linker fixed.

Base Model	Wikipedia Test		NYT Test
Base Model	Original	+MACE	Original	+MACE
GENRE	161325	107019	6927	4595
EveLink	24614	17501	2352	1731
SEFEL	446	446	59	59

Table 3: Efficiency study results. Runtime is measured in seconds. The reported values are downstream linker runtime with pre-cached MACE outputs; online MACE latency is analyzed in the text and Appendix A .

Stage	Evidence state/action
Raw role outputs	Time: “1935,” “7 April 1939”; participant: “Hughes,” “Billy Hughes”; additional cues: “UAP,” “Australia,” and “Australia and the War Today.”
Verified/filtered evidence	Removes the unrelated date “7 April 1939” and normalizes the overlapping “Hughes” mentions.
Coordinator action	REFINE , because the provisional candidate feedback remains broad.
Final retained evidence	“1935,” “Billy Hughes,” and “Australia and the War Today.”
Candidate change	The original list ranks Battle for Australia first and omits World War II from its 30 slots; MACE ranks World War II first.

Table 4: Case of evidence refinement for the “coming war” mention.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Acquisition setting	$K$	Wikipedia ACC (%)
EveLink	30	79.22
EveLink	50	80.27
EveLink	100	82.87
EveLink + MACE	30	84.80
SEFEL	30	83.64
SEFEL	50	83.68

Appendix

Table 5: Naive candidate-set expansion control.

MACE stage	Wikipedia (s)	NYT (s)
Time/location/participant/type extraction	2.67	2.46
Contextual verification	3.16	2.91
Deterministic filtering and lookup	0.12	0.10
First coordinator decision	1.75	1.61
Conditional refinement	1.24	0.29
Second coordinator decision	1.16	0.31

Appendix

Table 6: Engineering estimate of online MACE latency per mention, reconstructed from 200 samples per dataset with LLaMA3.1-8B and a two-round coordinator limit.

Role configuration	ACC (%)	Recall@5 (%)
Four roles (default)	86.67	93.33
Four roles + Instrument + Purpose + Cause	86.67	93.33

Appendix

Table 7: Role extensibility on a randomly sampled 5% of the Wikipedia test set using the same SEFEL checkpoint.

Method	Wikipedia			NYT
Method	All	Verb	Noun	All	Verb	Noun
SEFEL_Uni+Tag	83.28	85.27	80.85	62.43	66.72	49.66
SEFEL	83.64	85.99	81.29	76.90	79.61	68.85
SEFEL+MACE	84.83	87.00	82.67	79.44	81.41	73.59

Appendix

Table 8: Indirect comparison of traditional extraction, the original SEFEL front end, and MACE under the same downstream linker.

Stage	Traditional UniST + TagPrime	LLM-based SEFEL front end
Extracted evidence	Long span “Weakness in the stock market … President Clinton”; “to 116 42 yen” (duplicated); “the dollar”	Participant: Clinton; type: “impeachment proceedings”
Candidates induced by the extracted evidence	NIL; Conference proceeding; Dollar; Shadrake v Attorney-General; Proceedings (magazine)	Impeachment of Bill Clinton; Shadrake v Attorney-General; Impeachment inquiry against Donald Trump
Gold event present in this evidence-derived candidate list	No	Yes; it is the highest-scored entry in the stored list

Appendix

Table 9: Case study of evidence quality for the NYT mention “proceedings.”

Item	Value
arguments	Extraction, verification, and refinement modules return a JSON list of spans with type , span_text , start , ln , and confidence .
action	The coordinator returns one of accept , refine , reverify , drop , or rerun .
temperature	All LLM modules use 0.0 .
max_rounds	The coordinator loop uses at most 2 rounds.
lookup_topk	Candidate lookup keeps the top 10 titles for each provisional query.
L, U	Argument selection keeps at most L=5 arguments total and U=2 arguments per evidence type.

Appendix

Table 10: Output format and reproducibility parameters used by MACE.

ECHO: Event-Centric Hypergraph Operations via Multi-Agent Collaboration for Multimedia Event Extraction
Mar 4, 2026Hailong Chu, Hongbing Li, Yunlong Chu +6Event Extraction Multimodal Grounding
Beijing University of Posts and Telecommunications Beijing, China · Tianjin University Tianjin, China · Chongqing University of Posts and Telecommunications Chongqing, China +1
EnsembleLink: Accurate Record Linkage Without Training Data
Jan 29, 2026Noah DasanaikeEntity Resolution Computational Social Science
PhD Candidate, Department of Government, Harvard University
TimelineAgent: An LLM-based Framework for Query-Centric Event Timeline Extraction Across Multiple Documents
Oct 19, 2026Qingkai Min, Yunxin Bao, Yifeng Lu +4Event Extraction LLM Information Extraction
Zhejiang University · School of Engineering, Westlake University · Institute of Advanced Technology, Westlake Institute for Advanced Study +3

Mar 4, 2026J/K move · Enter open · S save

ECHO: Event-Centric Hypergraph Operations via Multi-Agent Collaboration for Multimedia Event Extraction

Hailong Chu, Hongbing Li, Yunlong Chu, Shutai Huang+5

Beijing University of Posts and Telecommunications Beijing, China · Tianjin University Tianjin, China · Chongqing University of Posts and Telecommunications Chongqing, China+1

Event Extraction Multimodal Grounding

Open paper AskPDF arXiv

Multimedia event extraction (M2E2) aims to predict triggers, ground arguments across text and images, and then assemble them into schema-consistent event records. Recent LLM-based approaches have shown strong potential for M2E2, but their intermediate event hypotheses often remain implicit, and event-argument linking is still tightly coupled with role binding. This leaves little opportunity to inspect or revise intermediate event hypotheses and makes predictions brittle to early errors. To bridge this gap, we present ECHO, a multi-agent framework that reframes M2E2 as iterative refinement over an explicit Multimedia Event Hypergraph (MEHG). Instead of relying on implicit linear generation, ECHO performs auditable atomic updates over a shared hypergraph, making intermediate event structures explicit and revisable. Furthermore, we introduce a Link-then-Bind strategy that decouples event-argument linking from role binding, reducing premature semantic commitment during structured prediction. Extensive experiments on the M2E2 benchmark show that ECHO consistently outperforms prior state-of-the-art approaches, achieving gains of 7.3 and 15.5 F1 points on event mention and argument role, respectively.

Abstract

Explore similar work

ECHO: Event-Centric Hypergraph Operations via Multi-Agent Collaboration for Multimedia Event Extraction

EnsembleLink: Accurate Record Linkage Without Training Data

TimelineAgent: An LLM-based Framework for Query-Centric Event Timeline Extraction Across Multiple Documents