Suppose a system groups news mentions into clusters that refer to the same real-world event ("the acquisition," "the deal," "bought") and the same participants. Good cross-document systems let the two tasks inform each other: knowing that two "acquire" mentions share a buyer makes them more likely to be the same event, and knowing they are the same event makes their buyers more likely to be the same company. Today this back-and-forth is an unguarded loop. Nothing guarantees that it settles, that its answer is independent of the order in which documents arrived, or that adding one new article leaves earlier decisions mostly intact. We propose to rebuild the joint event–entity loop as a contraction mapping. A contraction is an update that always pulls any two states closer together by a fixed factor below one. By a classical result (Banach's fixed-point theorem), such an update has exactly one resting state, reaches it from any starting point, and moves that resting state only a bounded amount when the input changes.
The main claim is about stability, not about beating the leaderboard. A contractive joint coreference model should match the accuracy of the best existing joint system. It should also give three properties current systems lack: output that is provably independent of iteration schedule and document order, a computable bound on how much existing clusters can change when a new document arrives, and cheap warm-started updates for a growing collection. These properties matter for any deployed event knowledge base, where analysts see clusters change under them. The file contains no prior work that measures any of them.
Core Claim
In the current state of the art, joint reasoning between events and their arguments either runs as a heuristic alternating loop or is avoided altogether. Revisiting Joint Modeling of Cross-document Entity and Event Coreference Resolution alternates agglomerative clustering of entities and events. In each round, it rebuilds event representations by averaging the current entity clusters of each event's who/what/where/when arguments. It reports 79.5 CoNLL F1 for events and 71.2 for entities on ECB+, but has no halting guarantee, and an early wrong merge permanently corrupts later rounds. Pairwise Representation Learning for Event Coreference avoids iteration entirely. It uses a RoBERTa cross-encoder over sentence pairs with separate argument-matching modules and reaches 84.4 CoNLL F1. Its remaining errors are dominated by cases where the evidence lies outside the sentence pair.
Separately, the fixed-point literature shows how to make learned iterations provably convergent. Plug-and-Play Methods Provably Converge with Properly Trained Denoisers does this with exact spectral normalization of convolutions. Implicit Graph Neural Networks does it with a Perron-Frobenius bound on graph message passing. Convergent Graph Solvers uses input-generated contracting linear maps. Deep Equilibrium Architectures for Inverse Problems in Imaging trains through the fixed point with implicit differentiation. These papers supply our method. None of them touches predicate–argument structure or coreference.
The property we exploit is the standard sensitivity bound for contractions. Let $F(x; u)$ be the joint update, where $x$ stacks the states of all event and entity mentions and $u$ is the input (the mention encodings and argument links). If $F$ is an $L$-contraction in $x$ with $L<1$, and is $M$-Lipschitz in $u$, then the unique fixed points for two inputs satisfy
$$\|x^*(u) - x^*(u')\| \le \frac{M}{1-L}\,\|u - u'\|.$$
This follows by applying the triangle inequality to $x^* = F(x^*; u)$. Adding a document changes $u$ only in the new mentions and their links. The bound therefore limits how far existing mention states can drift, and warm-starting from the old fixed point converges in roughly $\log(\text{tolerance}/\text{drift})/\log L$ steps. The novelty is to build this guarantee into structured event coreference and to test whether it survives realistic data.
Research Program
The model keeps a continuous state vector for every event mention and every entity mention in a topic. Mentions are gold, following the ECB+ protocol of the Barhom et al. paper. Each mention first receives a fixed base encoding from a RoBERTa encoder over its surrounding context. Argument links come from an off-the-shelf semantic role labeller, which marks each entity as the agent, patient, location, or time of an event, as in the pairwise paper.
One update of the state does two things. First, each event's state is recomputed from its base encoding plus role-specific linear messages from the current states of its argument entities. Each entity's state is likewise recomputed from its base encoding plus messages from the events it participates in, so information flows in both directions. Second, each mention mixes in a similarity-weighted average of same-type mentions. This is a soft stand-in for "the cluster I currently belong to," which Barhom et al. compute with hard clustering. The mixing weights are a bounded kernel of the current states, clipped to keep their Lipschitz constant known.
Contraction is enforced by construction rather than by a training penalty. The role-specific message matrices are spectrally normalized, and the argument adjacency is row-normalized. A global mixing coefficient $\gamma$ scales the whole state-dependent part so that the composite Lipschitz constant is provably below a chosen $L$. We prove this bound for the soft-clustering step explicitly; this step is the one genuinely new piece of analysis, because a state-dependent similarity kernel is not covered by the linear graph-solver results.
Inference solves for the fixed point with Anderson acceleration, a standard method for speeding up fixed-point iteration, and stops at a residual tolerance. Training uses implicit differentiation at the fixed point, as in the deep equilibrium imaging work, so memory does not grow with iteration count. The loss combines a pairwise coreference loss on the equilibrium states with the attractive and repulsive terms of Resolving Event Coreference with Supervised Representation Learning and Clustering-Oriented Regularization. Final clusters come from agglomerative clustering on equilibrium pair scores. The discrete step is a deterministic function of a unique fixed point, so the output no longer depends on the iteration schedule or on document order.
The known cost is a "performance tax": a contraction constraint can reduce accuracy. We characterize it by sweeping $L$. We also test a weaker condition, pseudo-contraction paired with a multi-step iteration, following Deliang Wei et al.. That work reports that this weaker condition preserves convergence while reducing the accuracy drop from strict contraction.
The streaming setting is where the guarantee should pay off. Within each ECB+ topic, documents arrive one at a time in random order. After each arrival the system updates its clusters, and we measure accuracy, churn, and compute at every step.
Baselines
The naive reference point is an append-only streaming system. It assigns new mentions to existing clusters using the pairwise scorer and never revises earlier decisions. Its churn is zero by construction, so it marks the accuracy cost of perfect stability. Beating it means reaching clearly higher final CoNLL F1 while keeping churn in the same low range.
The strongest accuracy competitor is PAIRWISERL from Paired Representation Learning for Event and Entity Coreference, which reports 84.4 CoNLL F1 for ECB+ events, 81.5 for entities, and 57.12 AVG-F on KBP 2017. We will retrain it with the same RoBERTa backbone, the same semantic role labeller, and the same hyperparameter search budget as our model. In streaming mode, we score each new document's mentions against all prior mentions and re-cluster from scratch at every step. On batch accuracy, we do not expect to beat it. Success there means non-inferiority within about one CoNLL point. The claim to beat it is on churn and on compute per added document, since its quadratic cross-encoder scoring must be rerun at each step while our base encodings are computed once.
The most direct competitor is the alternating joint model of Barhom et al., which reports 79.5 event and 71.2 entity CoNLL F1 on ECB+. To isolate architecture from encoder quality, we will re-implement it on the same RoBERTa encodings. We run it three ways: for its default number of rounds, to its empirical stopping point, and for ten times as many rounds. Beating it means equal or better F1, lower variance across document orders, and no degradation under extended iteration.
The key ablation removes the contraction constraint while keeping everything else fixed. It runs the same model unconstrained, either unrolled for $K$ steps or run to an equilibrium it is not guaranteed to reach. A second ablation replaces the hard constraint with a soft penalty on the step size $\|x_{k+1}-x_k\|$. If either ablation matches the constrained model on churn and order sensitivity, the guarantee is not doing the work.
For every streaming metric, no reported result was found on Lacuna for streaming or incremental cross-document coreference stability. Those numbers will be established here.
Why not just…?
Why not just use PAIRWISERL and re-cluster? On batch accuracy it is the model to beat, and we concede that. The file's own evidence shows how little joint structure adds to it: pairwise encoding gave 2.7 points, while structured arguments added only 0.4. Re-clustering from scratch after each new document, however, offers no bound on how many earlier links flip, and its cost grows with every pair of mentions. The streaming experiment measures both directly. If PAIRWISERL's churn turns out to be negligible in practice, our stability claim fails and we will report it.
Why not just run the Barhom et al. loop until it stops changing? This is the most serious threat. Their joint-over-disjoint gain was only about one point, which suggests the loop may already settle quickly on ECB+. If it converges, is insensitive to order, and does not drift under extended iteration, then a guarantee adds little in-distribution. We test this before building anything by instrumenting the loop on shuffled document orders and at ten times the usual iteration count. The divergence results of Simon Frieder et al. and Ali Joundi concern other iterative models. They motivate the concern but do not prove that this loop fails. If it is empirically stable, we would narrow the claim to the warm-start compute bound and to robustness when the input shifts, such as KBP news and discussion-forum text.
Why not just add a soft stability penalty or unroll a few steps? Unrolled models trained for a fixed depth show "peak-and-crash" collapse when run longer (Deep Equilibrium Architectures for Inverse Problems in Imaging). Contractive graph solvers kept roughly 1.13% error on 100-node graph value iteration where the less constrained implicit graph network reached roughly 29% (Convergent Graph Solvers). Both results come from other domains, so our key ablation compares soft penalty, unrolling, and the hard constraint directly on coreference.
Why not just use a long-context encoder to remove the sentence-pair bottleneck? A long-context encoder would help with evidence that lies outside the sentence pair, but it addresses a different problem. It still gives no guarantee about order or churn. It also cannot see a whole cross-document topic at once, and it does not address incremental updates.
Proposed Experiments
All experiments use ECB+ with gold mentions for cross-document event and entity coreference, and KBP 2015–2017 for within-document transfer. The accuracy metric is CoNLL F1, with MUC reported alongside. The stability metrics are churn, the standard deviation of CoNLL F1 across 20 random document orders per topic, and the pairwise disagreement in cluster assignments between orders. Churn is the fraction of previously decided mention-pair links that flip after each new document.
The first check is a pilot with no new model. We instrument the re-implemented Barhom et al. loop and PAIRWISERL re-clustering for churn and order sensitivity. If both are already near zero, we stop and redirect the project as described above.
The second check measures the stability–accuracy tradeoff. We sweep the contraction bound $L$ from 0.5 to 0.99, alongside the pseudo-contractive variant, and plot final CoNLL F1 against churn and order variance. The predicted outcome is a region where F1 is within one point of the unconstrained ablation while churn and order variance fall substantially.
The third check tests whether the theorem is informative. We compare the bound $M/(1-L)$ with the observed drift of old mention states after each insertion, and compare iterations needed with a warm start versus a cold start. A bound that is off by orders of magnitude would be technically true but useless, and we would say so.
The fourth check is a stress test. We run every model at ten times its trained iteration count, and apply the ECB+-trained models to KBP without retraining. A contractive model should keep both accuracy and fixed-point residual stable, while unconstrained variants may drift.
The decisive result is the combination of three findings. The constrained model is within about one CoNLL point of PAIRWISERL on ECB+ events. It shows markedly lower churn and order variance than both PAIRWISERL re-clustering and the Barhom et al. loop. And the soft-penalty ablation does not recover that stability.
References
- Revisiting Joint Modeling of Cross-document Entity and Event Coreference Resolution
- Pairwise Representation Learning for Event Coreference
- Plug-and-Play Methods Provably Converge with Properly Trained Denoisers
- Implicit Graph Neural Networks
- Convergent Graph Solvers
- Deep Equilibrium Architectures for Inverse Problems in Imaging
- Resolving Event Coreference with Supervised Representation Learning and Clustering-Oriented Regularization
- Learning Pseudo-Contractive Denoisers for Inverse Problems
- Paired Representation Learning for Event and Entity Coreference
- (Non-)Convergence Results for Predictive Coding Networks
- From sparse recovery to plug-and-play priors, understanding trade-offs for stable recovery with generalized projected gradient descent