Background Mentions Point to Anchors: Centrality-Shaped Cross-Document Event Coreference

Cross-document event coreference asks whether two event mentions in different texts, such as "the final" in one match report and "Sunday's win over the Netherlands" in another, describe the same real-world event. Most systems first group documents by topic, a step called preclustering, and then compare mentions only inside each group. On corpora where events are shared across different stories, this first step is disastrous. Applying it to the cross-subtopic football corpus FCC-T drops the LEA F1 score from 67.6 to 13.9, as reported in Cross-Document Event Coreference Resolution Beyond Corpus-Tailored Systems. Our proposal starts from a simple observation about news writing. When two articles about different stories share an event, the event is usually the main story of one article and a piece of background in the other. A report on a semifinal mentions the group-stage match in a "previously" paragraph, and that group-stage match is the central event of an earlier report. Comparing whole documents misses this link because the two documents are mostly about different things. Comparing a background mention against the other document's central event would catch it.

We propose to make this asymmetry the organizing structure of cross-document coreference. Each document gets an anchor, its central event, found with a sentence-level discourse profiler that does not depend on upstream coreference. Every non-anchor mention gets a functional role, such as Previous Event, Historical, or Anecdotal. Candidate generation and clustering then follow a hub-and-spoke pattern. Background mentions are retrieved against the anchors of other documents. Two background mentions merge only through a shared anchor or through strong direct evidence. The claim is that this recovers the cross-subtopic links that preclustering destroys, and that it cuts the false merges caused by long chains of lookalike peripheral mentions. Both effects should be largest exactly where current systems fail: FCC-T, GVC, and the densely annotated CDEC-WN.

Core Claim and What Is New

The claim is empirical first and methodological second. Empirically, we hypothesize that a large majority of gold cross-subtopic coreference links connect a mention in a Main Event or Consequence sentence of one document to a mention in a background-role sentence of the other. Links between two background mentions should be rare, and they should be disproportionately often the ones annotated as partial identity. Methodologically, if this holds, a role-conditioned anchor index should replace document preclustering. It should beat both preclustering and role-blind mention retrieval on cross-subtopic recall, at a candidate budget far below exhaustive pairwise comparison.

Prior work does not test this. Identifying the Most Dominant Event in a News Article by Mining Event Coreference Relations ranks events by the size, spread, and position of within-document coreference chains. It works only inside one document, and it collapses from about 72 to about 47 F1 on RED when chains are predicted rather than gold. That collapse motivates our choice to take centrality from discourse roles instead of from coreference chains. Discourse as a Function of Event supplies the method. It defines the eight functional roles around a Main Event and shows they help within-document coreference on KBP. It never asks how roles pair up across documents. CHEER uses predicted central events to prune candidate causal links inside a document, with central-event prediction near 80% F1 and an average gain of 5.9% F1. We borrow its pruning idea and move it across document boundaries and from causality to identity.

Two cross-document systems already use discourse structure, and they are the closest neighbors. Cross-Document Event Coreference Resolution on Discourse Structure compresses each mention pair into a T5-generated pseudo-document. It then parses rhetorical structure, meaning the tree of relations like Elaboration and Cause between clauses, and reaches 86.4 CoNLL F1 on ECB+. DIE-EC feeds within-document rhetorical trees and cross-document lexical chains into a graph network. Both treat discourse as a per-pair feature over generic rhetorical relations, and both are evaluated mainly on topic-isolated benchmarks. Neither uses the main-event hierarchy of each document to decide which pairs are compared at all, or how clusters may grow.

Research Program

The work has three parts. The first is a measurement study. We run the functional discourse profiler over FCC-T, GVC, ECB+, and CDEC-WN. To check the profiler's accuracy on new domains, we hand-label roles on a small stratified sample per corpus; football and gun-violence reports differ from the training news. We then tabulate role pairs over gold cross-document links. Links are split into within-subtopic and cross-subtopic on FCC-T, using the subtopic structure from Breaking the Subtopic Barrier, and into full and partial identity on CDEC-WN, using the annotations of Cross-document Event Identity via Dense Annotation. This study decides whether the method is worth building. If the anchor–background pattern holds for fewer than roughly two thirds of cross-subtopic links, the hub-and-spoke constraint is wrong and we stop.

The second part is the anchor index. Each document's anchor is the event mention set with the highest Main Event score from the profiler. Spread across sentences is added as a tiebreak, which avoids leaning on headline position; that matters because the inverted-pyramid assumption fails on chronological match reports. A bi-encoder in the style of Cross-document Event Coreference Search, which reports MRR@10 of 90.06 on CoreSearch, embeds marked mentions. The encoder gets a learned role embedding, and training uses role-asymmetric hard negatives. In this setup, a background mention queries the index of anchors and anchor-cluster members. An anchor mention queries other anchors and the background mentions that retrieved it.

The third part is hub-and-spoke clustering. A cross-encoder scores the retrieved pairs. Clustering proceeds in two stages. Anchors are first merged with each other under a strict threshold. Each background mention is then attached to its best anchor cluster. A background–background edge can create or join a cluster only above a higher, separately tuned threshold. This keeps two peripheral mentions from being chained together through weak similarity, which is the transitivity failure documented on CDEC-WN. We also report a soft variant in which background mentions may attach to an anchor as a subevent rather than as an identical event. This variant is scored against CDEC-WN's partial-identity links, but we treat it as secondary.

Baselines

The naive reference points are two. The first is lemma matching, which Cross-document Event Identity via Dense Annotation reports at 48.2 F1 on CDEC-WN. The second is the same neural pipeline we use, run with TF-IDF document preclustering. On FCC-T, that configuration's reported LEA F1 of 13.9 (Beyond Corpus-Tailored Systems) is the failure we aim to fix.

The strongest competitor on the cross-subtopic question is the neural system run without preclustering. It compares all mention pairs across the corpus and reaches 67.6 LEA F1 on FCC-T in the same paper. It is strongest because it never discards a candidate, so any recall our anchor index loses shows up directly against it. We will rerun it with our exact encoder (RoBERTa-large cross-encoder), training data, and hyperparameter search budget. Beating it means matching or exceeding 67.6 LEA F1 on FCC-T while scoring at least ten times fewer candidate pairs. A clear win would be gaining at least two LEA F1 points through fewer false merges.

On ECB+, the reference numbers are 79.5 CoNLL F1 for the joint entity–event model of Revisiting Joint Modeling and 86.4 for the discourse-plus-Longformer system of Cross-Document Event Coreference Resolution on Discourse Structure. We do not expect large gains on ECB+, because the benchmark has only about ten cross-subtopic clusters and our mechanism has little to act on there. The bar on ECB+ is no loss beyond one CoNLL point. On CDEC-WN, the baseline is the BERT cross-encoder at 57.6 F1 from Cross-document Event Identity via Dense Annotation. On GVC, we found no reported result on Lacuna for an absolute LEA or CoNLL score, so we will report our rerun baselines only.

The key ablation is role-blind retrieval. It uses the same bi-encoder, the same candidate budget, and the same clustering thresholds, with role embeddings and asymmetric negatives removed and with ordinary agglomerative clustering. This isolates whether centrality, rather than mention retrieval itself, does the work. Two further ablations swap the profiler-based anchor for chain-based centrality computed from predicted within-document coreference, and for a headline-trigger anchor.

Why not just…?

Why not just use the headline event as each document's anchor? That anchor is cheap. But Identifying the Most Dominant Event reports a headline baseline at about 45 F1 for finding the dominant event, against about 72 for chain mining with gold coreference. Match reports often lead with a hook, such as a red card or a record, rather than the match itself. The headline-anchor ablation measures how much of our coreference gain survives with this weaker anchor. If it survives nearly intact, we will say so and recommend the simpler anchor.

Why not just retrieve candidates at the mention level, as in coreference search, and skip centrality entirely? This is the most serious threat. Mention-level dense retrieval already avoids document preclustering, and Cross-document Event Coreference Search shows it works at scale on Wikipedia. Our bet is that role-blind retrieval spends its budget on lookalike background mentions. On GVC, for example, there are many similar shootings that differ mainly in date, and Beyond Corpus-Tailored Systems found that masking time costs 23.5 LEA points there. The role-blind ablation settles this directly. If it matches our system on cross-subtopic recall and LEA at equal budget, centrality adds nothing, and the honest conclusion is a negative one.

Why not just compare all pairs with a strong encoder, or a fine-tuned large language model? Exhaustive comparison is the strongest baseline above, and we beat it only if constrained clustering removes false merges, not merely by saving compute. Fine-tuned LLaMA-2-7B outperformed a modular pipeline on cross-document event extraction in Harvesting Events from Multiple Sources, at 79.1% against 72.6% F1. That paper also notes that exhaustive cross-document comparison by such models does not scale. We will include a fine-tuned 7B pair scorer on the FCC-T test set, run under both candidate sets, which separates scorer quality from candidate structure.

Decisive Checks

The first check is the measurement study. Gold cross-subtopic links on FCC-T must be dominated by anchor–background pairs, and background–background links must be enriched for partial identity on CDEC-WN. Otherwise the program stops. The second check is FCC-T, run on both training and test data with the setup above. There, the anchor index must reach at least 95% of the cross-subtopic link recall of exhaustive comparison while scoring at least ten times fewer pairs, and it must exceed role-blind retrieval at the same budget. The third check is whether hub-and-spoke clustering raises precision on CDEC-WN and GVC over agglomerative clustering on identical scores. Precision is measured by LEA and by the fraction of predicted merges between two background mentions that are wrong. Together these show whether narrative centrality helps decide which mentions to compare and how clusters may grow, as opposed to serving as one more pairwise feature.