Counting Rivals, Not Sentences: Intervening-Competitor Recency for Event Coreference

Event coreference resolution groups the words in a text that describe the same real-world happening, such as "exploded," "the blast," and "this." Systems that do this well within a single sentence break down when the mentions are far apart. On the TAC-KBP 2017 news-and-forum benchmark, a strong document-level model reaches 51.2 average F1 overall but only 37.1 F1 on mention pairs separated by more than three sentences (Improving Event Coreference Resolution Using Document-level and Topic-level Information). The usual reading is that distance is the problem. We propose a different reading. A distant antecedent is hard to find mainly when other plausible antecedents appear in between, and the length of the gap matters much less. A "blast" mentioned five sentences after the only explosion in the article should be easy. A "blast" mentioned two sentences after a second, unrelated bombing should be hard.

The proposal makes that idea measurable and then builds it into a model. We redefine recency for events as the number of intervening rival mentions: mentions between the candidate antecedent and the anaphor that look like the same kind of event. We first test whether this rival count explains long-distance errors better than sentence distance does. If it does, we add a learned penalty on the rival count to a modern span-pair scorer, and let the shape of the penalty depend on the form of the referring expression (verb, event noun phrase, or demonstrative). The contribution is a falsifiable claim about what "recency" means for event anaphora. That claim is tested by a diagnostic that comes before any modeling, with a stated condition under which we abandon the idea.

Core Claim

Recency matters for antecedent choice because the most recent compatible mention tends to win the competition for the listener's attention, and time passing is a weak proxy for that. Event coreference is where this distinction should matter most. Event anaphors sit about three sentences from their verbal antecedents on average (Resolving Event Noun Phrases to Their Verbal Mentions), which is farther than entity pronouns, about 71% of which resolve within their own sentence (Analysis of Syntax-Based Pronoun Resolution Methods). News reports also often describe several events of the same type. We claim that, once the model's semantic score is held fixed, the probability that a candidate is the correct antecedent falls with the number of intervening same-type rivals, and falls much less with raw sentence distance. We further claim that a scorer which penalizes rivals recovers most of the long-distance gap without the cost of discourse parsing.

Relation to Prior Work

Three earlier papers already treat distance as a signal for event linking, and they directly motivate this work. A Hierarchical Distance-dependent Bayesian Model for Event Coreference Resolution clusters event mentions with a prior that decays with a learned pairwise distance. That distance is a feature-based similarity and does not count intervening mentions, and inference by sampling scales poorly. Resolving Event Noun Phrases to Their Verbal Mentions ranks verbal antecedents for event noun phrases with parse-tree kernels and finds positional features highly discriminative. It covers only noun-to-verb links and depends on gold parses. The closest prior idea comes from Graph Based Decoding for Event Sequencing and Coreference Resolution. Its authors state an "explain away" intuition: a link to an earlier mention is less likely if similar mentions intervene. They enforce it through hand-built features and transitive reduction inside a perceptron decoder, and they report sentence distance among the most important features. We take that intuition and make it the definition of recency. We do this in three ways their work does not: we test it as an explanatory variable against sentence distance, we compute the rival count softly from the neural model's own similarity judgments, and we condition its effect on the form of the anaphor.

Two results put limits on the design. Event versus entity co-reference: Effects of context and form of referring expression shows in human experiments and on the ParCorFull corpus that "this" favors event antecedents while "it" favors entities, and that a recently active event raises the chance of later event reference. This is why the rival penalty must vary with anaphor form and cannot be a single curve. A Context-Dependent Gated Module for Incorporating Symbolic Semantics into Event Coreference Resolution finds that simply concatenating noisy symbolic features can hurt, and that gating combined with noise-aware training is needed. The rival count depends on predicted event types, so we feed it in through a soft, gated penalty and avoid a hard feature for the same reason.

Two discourse-structure papers, Cross-Document Event Coreference Resolution on Discourse Structure and Enhancing Cross-Document Event Coreference Resolution by Discourse Structure and Semantic Information, take a competing view. They replace linear distance with paths through rhetorical-structure trees, which are parses of how clauses support one another, and they gain substantially on ECB+. Our hypothesis offers a cheaper explanation for part of that gain. A distant nucleus clause may stay reachable partly because few rival events appear in the satellite clauses that separate it from the anaphor. We test this directly instead of assuming it. The modern span-based scorers we build on supply the method: Span-Based Event Coreference Resolution and Learning Event-aware Measures for Event Coreference Resolution.

Research Program

The first phase is diagnostic, and every later phase depends on it. We reproduce a document-level event coreference model on KBP 2015/2016 training data and KBP 2017 test data, following the split in Wu et al. For every gold anaphor and every candidate antecedent we record three quantities: the sentence distance, the number of intervening gold mentions of the same event subtype, and the model's raw pair score. We then fit a logistic regression that predicts whether a pair is a true link from the raw score plus either distance, rival count, or both.

The idea survives only if two conditions hold. Rival count must add predictive value over sentence distance, measured by held-out log-likelihood with bootstrap intervals. Long-distance pairs with zero rivals must also be resolved substantially better than long-distance pairs with rivals. If sentence distance absorbs the rival effect, we stop: an ordinary distance feature would then do the job, and the proposal reduces to an existing practice. We run the same analysis on the OntoNotes noun-to-verb subset used by Chen et al., which checks the claim on links across parts of speech.

The second phase builds the model. For an anaphor $i$ and candidate antecedent $j$, the score is

$$s(i,j) = f(m_i, m_j) - \lambda_{\phi(i)}\, g\big(c_{ij}\big), \qquad c_{ij} = \sum_{j<k<i} \sigma\big(f(m_k, m_j)\big),$$

The symbols are as follows. $f$ is the base pairwise scorer applied to span representations $m$. $\sigma$ is the logistic function, so $c_{ij}$ is a soft count of intervening mentions $k$ that the model itself judges similar to the candidate $j$. $g$ is a small monotone network, kept nondecreasing by using positive weights. $\phi(i)$ is the form class of the anaphor (verbal trigger, event noun phrase, or demonstrative pronoun), and $\lambda_{\phi(i)}$ is a learned weight for that class.

Because the count reuses $f$, the method does not depend on an external event-type classifier. Following Lai et al., we also train a variant that adds predicted-subtype agreement to the count through a gate, with label noise injected during training. A third variant replaces the soft count with transitive-reduction decoding in the style of Liu et al. on top of the neural scorer. Comparing these tells us whether the explain-away effect is better placed in the scorer or in the decoder.

The third phase analyzes what was learned. We plot the learned penalty curves for each form class and compare them with the qualitative pattern Loáiciga et al. report for demonstratives versus other forms. We also measure how much of the published gain from discourse trees on ECB+ disappears once rival count is controlled.

Baselines

The naive reference point is a rule with no learning. It links each anaphor to the nearest preceding mention that has the same predicted event subtype, and starts a new cluster when there is none. This is the event analogue of the nearest-compatible-noun-phrase baseline that works well for noun phrase ellipsis (Using Syntax to Resolve NPE in English). No reported result for this rule on KBP 2017 was found on Lacuna, so we report it ourselves. Any model that fails to beat it on the long-distance subset has learned nothing beyond recency.

The strongest existing within-document competitor is the document-level and topic-level model of Improving Event Coreference Resolution Using Document-level and Topic-level Information. It reports 51.2 average F1 on KBP 2017 and 37.1 F1 on its category of pairs more than three sentences apart, compared with 32.1 for a chunk-based RoBERTa baseline. It is the right competitor because it already attacks long distance with full-document attention. That is exactly the "transformers make recency obsolete" position. We add our penalty to this model's scorer, keep its encoder, data split and trigger inputs unchanged, and give the original and the augmented version the same hyperparameter-search budget over five seeds. If our reproduction falls short of the published figures, we report both and compare against the reproduction.

We also report Span-Based Event Coreference Resolution, which reaches 47.1 average F1 on KBP 2017, as an end-to-end reference. On ACE 2005 the reference is Learning Event-aware Measures for Event Coreference Resolution, with 62.11 end-to-end F1 and 86.63 F1 with gold triggers. On the OntoNotes noun-to-verb task it is the twin-candidate tree-kernel model of Chen et al. at 61.36 F-score.

Beating the strongest baseline means three things. Long-distance F1 must improve by more than the variation across seeds. Overall average F1 must not fall. And the gain must exceed that of the sentence-distance ablation below.

That ablation isolates the key ingredient, so it matters more than any other comparison. In the distance-only control, the same penalty network takes bucketed sentence distance in place of the rival count, with identical capacity and tuning budget. A second ablation removes the form conditioning, so a single $\lambda$ is shared across all anaphor forms. A third replaces the soft count with a count based on gold subtypes, which gives an upper bound and shows how much is lost to noise in the soft count.

Why not just…?

Why not just add a sentence-distance embedding to the pair scorer? This is the most serious threat. Distance buckets are cheap and standard, and Liu et al. found sentence distance to be among the most important features. Rival count and distance are correlated, so a distance feature could capture most of the effect. The diagnostic in phase one exists to settle this question before any model is built, and the distance-only ablation settles it again under equal capacity. If the two perform within seed noise of each other, we will report that the rival hypothesis adds nothing.

Why not just restrict the search to the previous three sentences? The average gap of three sentences reported by Chen et al. is an average. Wu et al.'s long-distance category is defined as exactly the pairs that such a window would remove. A hard window therefore caps recall on the pairs we care about. We include the window as a decoding variant and report its recall ceiling on that category, which bounds what it can achieve.

Why not just rely on long-context attention, or use discourse trees? Wu et al. already use full-document Longformer attention and still score 37.1 on distant pairs, so attention alone has not closed the gap. Discourse trees do help on ECB+ (RoBERTa rises from 81.1 to 85.7 CoNLL F1 in Cross-Document Event Coreference Resolution on Discourse Structure). They do so at the cost of a frozen external parser. Our phase-three analysis tests whether that gain shrinks once rival count is controlled. If it does not shrink, the trees capture something that rival counting does not, and we will report that result.

Proposed Experiments

The decisive checks come in order, and each stage depends on the one before it. First, on KBP 2017 and the OntoNotes noun-to-verb links, rival count must predict true links beyond sentence distance and the base pair score. It must also separate long-distance pairs with no rivals from long-distance pairs with rivals, which is the predicted pattern, with zero-rival pairs resolved far more accurately. Second, the rival-penalized version of the Wu et al. model must improve long-distance F1 over both the original and the distance-only control, using the same five seeds and the same tuning budget. Overall average F1 is reported with MUC, B-cubed, CEAF-e and BLANC. Third, the gain must carry over to ACE 2005 with both gold and predicted triggers. With gold triggers the comparison is against 86.63, so any improvement cannot be blamed on trigger detection.

Fourth, for demonstrative event anaphors we evaluate on the ParCorFull English event and clausal links. No reported coreference F1 for this setting was found on Lacuna, so we compare only against our own baselines. There we check whether the learned penalty for "this" is flatter than the penalty for event noun phrases, as the form-bias findings of Loáiciga et al. predict.

A positive result would give event coreference a cheap, interpretable bias aimed at the long-distance failure. A negative result at the first check is still informative. It would show that the long-distance gap is a matter of semantic representation and not of competition between candidates, and future effort would then belong in the encoder.