The Linker Can’t Rank What It Never Sees

TL;DR for operators

When a system must map an ambiguous phrase in text to a canonical event in a knowledge base, enlarging the search pool is not necessarily the best use of compute. The more consequential failure may occur earlier: weak contextual evidence produces the wrong candidates, leaving the final linker nothing useful to rank.

MACE tests a different intervention. It inserts an LLM-based evidence-acquisition layer before an existing event linker. Specialized agents extract time, location, participant, and event-type clues; a verifier and deterministic filters remove weak evidence; a bounded coordinator can revise the evidence after inspecting provisional retrieval results; and a selector keeps only a small number of salient arguments. The resulting candidates are then passed to the existing linker without changing its architecture.

Across three linkers and two benchmarks, this upstream retrofit improves accuracy. More importantly for system design, controls that increase the candidate-set size without improving the evidence do not reach the same performance. The paper therefore supports spending retrieval budget on which evidence generates candidates, not only on how many candidates survive.

The trade-off is latency. Cached MACE outputs make the downstream stage look inexpensive, but reconstructed online processing adds an estimated 12.23 seconds per Wikipedia mention and 9.62 seconds per NYT mention. For batch enrichment this may be manageable; for interactive systems it is a substantial engineering constraint.

The failure can happen before ranking starts

Suppose a document contains a short phrase such as “coming war” or “proceedings.” A downstream system may have a strong scoring model, but that model can only choose among events presented to it. If the correct event never enters the candidate pool, better ranking cannot recover it.

This is the candidate-acquisition problem: deciding which knowledge-base events the final linker is allowed to consider. In Enhancing Event Candidate Acquisition for Event Linking, Zhang and colleagues introduce MACE as an upstream plug-in designed specifically for this stage.¹

The distinction matters operationally. MACE is not a replacement event linker. It changes the evidence used to retrieve candidate events and then hands those candidates to existing systems. The paper tests this arrangement with GENRE, EveLink, and SEFEL rather than building a new downstream architecture around the acquisition mechanism.

That separation lets the experiments ask a useful engineering question: if candidate generation improves while the linker stays fixed, how much accuracy can move?

Better evidence moves all three tested linkers

The main benchmark results answer that question consistently.

The magnitude varies sharply by backbone and dataset. GENRE gains 17.10 percentage points on Wikipedia and 19.24 points on NYT. EveLink gains 5.58 and 35.26 points. SEFEL, already stronger at baseline, gains 1.19 and 2.54 points.

For the smallest headline improvement, SEFEL on Wikipedia, the paper reports a paired-bootstrap 95% interval of +0.86 to +1.54 percentage points and an exact McNemar test with $p=3.65\times10^{-12}$. Comparable significance testing is not reported across every backbone-dataset combination, so the broader table should be read as comparative benchmark evidence rather than uniformly tested effect estimates.

Still, the cross-backbone pattern is relevant to retrofit decisions. The same acquisition layer improves three heterogeneous linkers without requiring their internal architectures to change.

More candidates do not reproduce the result

A natural explanation is that MACE simply retrieves more possibilities. The appendix directly tests that interpretation by increasing $K$, the number of retained candidates, without MACE.

For EveLink on Wikipedia, raising $K$ from 30 to 100 increases accuracy from 79.22% to 82.87%. MACE reaches 84.80% while still using $K=30$. SEFEL barely changes as $K$ grows: 83.64% at 30 candidates versus 83.70% at 100, compared with 84.83% for MACE at 30.

This candidate-expansion experiment is best read as a control on mechanism, not as a second headline benchmark. It weakens the explanation that the improvement comes from a wider search net alone.

The operational replacement is more specific: candidate quantity and candidate quality are different levers. If the acquisition query is built from noisy or generic evidence, allowing more results through can enlarge the wrong neighborhood of the knowledge base. MACE instead tries to improve the information entering retrieval.

Candidate acquisition becomes a staged evidence pipeline

MACE does more than ask one model to extract arguments.

Four specialized agents first look separately for time, location, participant, and event-type evidence. An LLM verifier, normalization, and deterministic filtering then remove unsupported, contradictory, duplicate, pronominal, or overly distant evidence.

Next, a coordinator gets provisional retrieval feedback and can choose among five actions: accept, refine, reverify, drop, or rerun. This loop is capped at two rounds. The coordinator is therefore not an open-ended agent conversation; it is a bounded correction mechanism tied to candidate lookup.

Finally, verified arguments are scored using extractor confidence, role prior, novelty, and proximity to the mention. MACE keeps at most five arguments overall and no more than two of each evidence type. Candidate queries are then constructed from the mention and selected evidence, merged, truncated to the top 30 events, and supplemented with NIL.

The ablations clarify why this staging exists. With SEFEL fixed downstream, full MACE reaches 84.83% on Wikipedia and 79.44% on NYT. Removing specialized agent extraction lowers those scores to 82.22% and 56.53%. Removing verification lowers them to 84.11% and 76.92%; removing the coordinator yields 83.76% and 75.97%.

Specialized extraction produces the largest measured contribution, particularly on NYT, while verification and revision provide smaller complementary gains. The evidence supports treating acquisition as a sequence of extraction, quality control, feedback, and selection rather than one generic extraction call.

The business case is retrofitability, not free inference

For teams operating knowledge-base linking pipelines, the paper suggests a practical diagnostic sequence.

If errors occur because good candidates are already retrieved but ranked poorly, MACE does not directly address the central problem. If the correct canonical event is frequently absent from the candidate set, however, replacing the entire linker may be premature. An upstream evidence layer can potentially improve what an existing linker gets to see.

That pattern extends conceptually to event-centric media intelligence, knowledge-base enrichment, and question-answering systems that depend on canonical event resolution. This is a Cognaptus inference from the system design, not something directly benchmarked in the paper. The affected decision is whether to invest first in a new ranker, a larger retrieval pool, or better evidence construction.

The deployment boundary is equally concrete. Runtime measurements showing GENRE and EveLink becoming faster with MACE use pre-cached MACE outputs, so they exclude the LLM acquisition stage. The paper separately reconstructs expected online MACE latency at 12.23 seconds per Wikipedia mention and 9.62 seconds per NYT mention.

That profile favors batch or precomputation-heavy workflows far more than latency-sensitive interactive ones. Selective invocation, faster inference, or additional serving work would be needed before treating the architecture as an online default.

What remains unproven

The evidence is strong for the tested comparative question: on Wikipedia and NYT, across GENRE, EveLink, and SEFEL, improving retrieval-oriented evidence acquisition can raise linking accuracy without modifying the downstream linker.

It is narrower than a general claim about agentic retrieval. The paper does not test additional languages, knowledge bases, broad domain shifts, or many other linker families. Online latency is reconstructed from 200 samples per dataset rather than presented as the primary end-to-end serving benchmark. Statistical testing is also concentrated on the SEFEL-Wikipedia comparison.

An exploratory role-extension test adds Instrument, Purpose, and Cause agents to the default four roles and finds no improvement in accuracy or Recall@5 on a random 5% Wikipedia sample. That result does not show additional roles are generally useless; it shows that more extraction roles are not automatically valuable. Evidence must earn its place by improving retrieval.

Fix the candidate set before rebuilding the judge

The most transferable idea in MACE is not that multiple agents improve event linking. It is that retrieval deserves its own evidence-quality pipeline.

The paper decomposes candidate acquisition into specialized extraction, verification, feedback-driven revision, salience selection, and bounded candidate construction. Its controls then show why that decomposition matters: retrieving a larger pool does not produce the same result.

For operators, the decision is consequently more precise. When the downstream model is choosing from poor candidates, improving the chooser may leave the actual bottleneck untouched. First inspect what evidence created the candidate set, whether that evidence is supported, and whether each retained clue meaningfully narrows the search.

That can preserve an existing linker architecture. It does not make the additional inference cost disappear.

Cognaptus: Automate the Present, Incubate the Future.