Enhancing Event Candidate Acquisition for Event Linking

1 Introduction

• Event linking associates text event mentions with knowledge-base entries or recognizes them as out-of-KB events, supporting text understanding and downstream applications.The paper also notes that event knowledge in text can enrich knowledge bases.Source: §1 · p. 1
• Existing approaches include retrieve-and-rank, generation-based, and fast argument-aware representation-based methods.Examples include BLINK-style retrieval and ranking, GENRE-style title generation, and SEFEL-style shared-space scoring.Source: §1 · p. 1
• Arguments differ in retrieval value, and unverified or peripheral spans can expand searches toward irrelevant candidates and weaken linking decisions.The example contrasts a participant that narrows candidates with a location that matches many events.Source: §1 · p. 1
• MACE is presented as a plug-in multi-agent pipeline that extracts, verifies, and revises event evidence before candidate construction, improving performance across different backbones without architecture changes.The paper describes this as the first plug-in for enhancing candidate event generation in event linking.Source: §1 · pp. 1–2

2 Task Definition

• Given text, a target event mention, its local context, and a knowledge base, event linking predicts an in-KB event label or NIL for an uncovered event instance.The candidate label space is the KB event entries together with NIL.Source: §2 · p. 2

3 Method

• MACE combines evidence extraction, verification, coordinator-guided revision, and candidate construction to refine event structure before linking.Figure 1 is identified as an overview of the framework.Source: §3 · p. 2, §3.1 · p. 2, §3.2 · p. 2

3.1 Agent-Based Argument Extraction

• Four evidence-specialized LLM agents focus on time, location, participant, or event-type arguments, producing typed raw evidence for a mention and its local context.Each argument records its evidence type, text span, normalized form, extractor confidence, and distance from the mention.Source: §3.1 · p. 2

3.2 Verification

• Verification checks arguments against the mention and context, merges duplicate forms, and removes pronominal, contradictory, or overly distant arguments.The result is a context-supported pool of typed arguments grounded in local context.Source: §3.2 · p. 2

3.3 Coordinator-Guided Evidence Revision

• A bounded coordinator loop observes evidence and lookup results, then accepts, refines, reverifies, drops, or reruns evidence before salience-aware selection.Selection ranks arguments using confidence, role prior, token novelty, and proximity, while removing duplicates and enforcing total and per-type caps.Source: §3.3 · pp. 2–3

3.4 Candidate Construction

• Candidate construction combines the mention alone with inputs containing retained arguments, submits each to the KB dictionary, and merges top-ranked results.The construction supports insertion before SEFEL, EveLink, and GENRE without modifying their architectures.Source: §3.4 · p. 3

4 Experiments

• Experiments use Wikipedia and NYT, with LLaMA3.1-8B for MACE and 30 candidate events; the reported evaluation includes in-KB and out-of-KB NYT mentions.Wikipedia has 19,267 test mentions; NYT has 769 in-KB and 993 out-of-KB test mentions.Source: §4 · p. 3

4.1 Effectiveness Study

• MACE consistently improves GENRE, EveLink, and SEFEL on the event-linking benchmarks, supporting its use as a plug-in across backbones.Table 1 reports accuracy; the SEFEL Wikipedia comparison is 83.64% versus 84.83%, a statistically supported 1.19 percentage-point gain.Source: Table 1 · p. 4, §4.1 · p. 4

4.2 Ablation Study

• Removing agent extraction, verification, or the coordinator lowers accuracy; agent extraction causes the largest drop, especially on NYT.The w/o Coordinator result isolates coordinator-guided revision while retaining the final salience selector.Source: Table 2 · p. 4, §4.2 · p. 4
• On Wikipedia, increasing EveLink from K = 30 to K = 100 improves accuracy by 3.65pp but remains 1.93pp below EveLink+MACE at K = 30.For SEFEL, the same expansion changes accuracy by 0.06pp, while MACE yields a 1.19pp gain at K = 30.Source: §4.2 · p. 4

4.3 Efficiency Study

• With cached MACE outputs, adding MACE reduces downstream runtime for GENRE and EveLink and leaves SEFEL unchanged.Table 3 compares each original model with the same model augmented by MACE under the same evaluation setting.Source: Table 3 · p. 4, §4.3 · p. 4
• Online reconstruction estimates an additional 12.23 seconds per Wikipedia mention and 9.62 seconds per NYT mention; batch preprocessing can amortize this cost.Latency-sensitive applications may need caching, a faster backend, or selective MACE invocation for ambiguous mentions.Source: §4.3 · pp. 4–5

4.4 Qualitative Analysis

• The “coming war” case shows verification removing unsupported evidence and coordinator feedback producing a more retrieval-useful argument set.The observed gain comes from removing a misleading date, consolidating duplicate participants, and retaining evidence that yields a more useful candidate query.Source: Table 4 · p. 5, §4.4 · p. 5

• Event-linking systems commonly use retrieve-and-rank architectures, with EveLink adding named entities, ArgEvelink adding arguments, and SEFEL using an argument-aware bi-encoder.MACE is orthogonal to these frameworks because it improves candidate acquisition before the event-linking model receives candidates.Source: §5 · p. 5
• MACE’s event-argument scope is narrower than general extraction: it seeks a small set of disambiguating arguments, not exhaustive annotation.It emphasizes evidence-specialized extraction, verification, and salience ranking before arguments are used for candidate acquisition.Source: §5 · p. 5
• MACE targets candidate event acquisition, which prior agent-based extraction work does not address for event linking.It uses specialized agents to acquire typed arguments, verify them, and retain those most useful for acquisition and linking.Source: §5 · p. 5

• MACE refines event structure before linking through candidate event acquisition.Source: §6 · p. 5
• Experiments show that MACE improves many event-linking models on both Wikipedia and NYT.Source: §6 · p. 5

• MACE’s reported efficiency comparison uses cached preprocessing, excluding the LLM calls needed to create its outputs.Fully online use introduces additional API latency and cost.Source: §8 · p. 6
• Transfer evaluation is limited mainly to SEFEL, EveLink, and GENRE on Wikipedia and NYT.The authors identify broader validation across domains, knowledge bases, and architectures as necessary.Source: §8 · p. 6
• MACE covers time, location, participant, and event-type evidence, while domains requiring finer-grained event semantics may need additional agents.Source: §8 · p. 6

A.1 Candidate-Set Expansion Control

• EveLink’s accuracy gain from increasing K from 30 to 100 is 3.65pp, yet it remains 1.93pp below EveLink+MACE at K = 30.For SEFEL, increasing K changes accuracy by 0.06pp, while MACE gives a 1.19pp gain at K = 30; the control table covers naive candidate-set expansion.Source: Table 5 · p. 7, §A.1 · p. 7

A.2 Latency Breakdown

• Online latency estimates include LLM calls and deterministic lookup stages excluded from cached downstream runtime, and do not quantify monetary API price.The table estimates per-mention latency from 200 samples per dataset using LLaMA3.1-8B and a two-round coordinator limit.Source: §A.2 · p. 7, Table 6 · p. 7

A.3 Role Extensibility

• The role-extensibility evaluation samples 5% of the Wikipedia test set and uses the same SEFEL checkpoint.Source: Table 7 · p. 7
• Additional roles can be inserted without changing the pipeline interface, but they do not improve the sampled subset.The evaluation uses the same SEFEL checkpoint on a randomly sampled 5% of the Wikipedia test set.Source: §A.3 · p. 7, Table 7 · p. 7
• Cause and Purpose evidence is often mapped to NIL as standalone entity spans, motivating the four broadly applicable default roles.Source: §A.3 · p. 7

A.4 Extraction-Front-End Comparison

• All settings use the same downstream SEFEL linker and differ only in the front end supplying evidence or candidates.Source: §A.4 · p. 7
• For the NYT mention “proceedings” in the sentence about impeachment proceedings against President Clinton, the gold event is Impeachment of Bill Clinton.Table 9 details how evidence and induced candidates differ between the traditional and LLM-based front ends.Source: §A.4 · p. 7
• The LLM-based front end recovers participant and event-type evidence that makes the gold event Impeachment of Bill Clinton retrievable.The traditional extractor's spans may be locally reasonable under its extraction formulation, but are less discriminative for event linking.Source: §A.4 · p. 7

B Prompt and Implementation Details

• The appendix includes the full prompt templates and decoding constraints used by MACE.It states that the code repository is available.Source: B Prompt and Implementation Details · p. 7

B.1 Prompt Inventory

• MACE’s reported prompt family includes TimeAgent, LocationAgent, ParticipantAgent, EventTypeAgent, VerifyAgent, RefineAgent, and CoordinatorAgent.Source: §B.1 · p. 7
• Extraction-style agents share a JSON contract, while the coordinator uses a separate action schema.Source: §B.1 · p. 7

Candidate Lookup Templates

• For each mention, lookup uses three query families: the mention alone, the mention paired with each retained argument, and the mention paired with the highest-scoring retained argument plus additional retained arguments.Source: Candidate Lookup Templates · p. 10
• Each query is submitted independently to the KB dictionary; duplicate returned events are merged before top-K truncation.Source: Candidate Lookup Templates · p. 10
• The final candidate set is augmented with NIL and passed unchanged to the downstream event linker.Source: Candidate Lookup Templates · p. 10