Modern event extractors read a document and fill in who did what to whom for every event it mentions. The strongest of them predict all events jointly, so that knowing one event shapes how the model reads the others. An Attack makes a Die event more likely, and a Buyer suggests a Seller is nearby. This joint prediction raises accuracy. It also opens a failure mode that the latent-variable literature knows well under another name. In a variational autoencoder, a strong decoder can explain its output without using the latent code, so the code collapses to an uninformative prior. This is called posterior collapse. We argue that a strong joint event decoder can do the same thing to the text. It fills a role because the surrounding events say that role usually appears, not because the sentence supports it. We call this structural collapse. When it happens, the model's confidence comes from the co-occurrence prior, so ordinary uncertainty measures do not notice it, and active learning built on those measures will not ask for labels where the prior is wrong.
The proposal has two parts, and both are cheap. First, we measure structural collapse directly. One trained table-filling extractor is run twice on the same document: once seeing all co-occurring events, and once seeing each event alone. We then ask, role slot by role slot, how far the joint view moves away from the evidence-only view. Second, we use the documents where the two views disagree most, especially where the joint view is confident, as the acquisition signal for active learning. Active learning here means the loop in which a model chooses which unlabeled examples a human should annotate next. Our prediction is concrete: under a fixed annotation budget, this "structural disagreement" acquisition reaches a given argument-classification F1 with fewer labeled documents than entropy, MC-dropout, or loss-prediction acquisition. The gain should be concentrated on rare roles and on documents whose event structure breaks the common co-occurrence patterns.
Core Claim
Joint dependency modeling and uncertainty-based active learning are in tension, and posterior collapse explains why. Generating Sentences from a Continuous Space showed that autoregressive decoders model local dependencies so well that a global latent code becomes redundant. Posterior Collapse and Latent Variable Non-identifiability proved that collapse happens exactly when the latent is non-identifiable, meaning the output distribution does not change with it. Replace "latent code" with "local textual evidence for a role" and "decoder" with "the cross-event attention of a joint extractor", and the same shortcut appears. Wherever the co-occurrence structure alone determines a prediction, the text's contribution to that prediction is not identifiable from the training signal.
A collapsed latent has a variance that says nothing about the input. In the same way, a structurally collapsed role slot has a confidence that says nothing about whether the text supports it. MC dropout makes this worse rather than better. Every dropout sample shares the same learned co-occurrence prior, so a prior-driven error is stable across samples and therefore looks certain.
The new element is that we do not try to read model uncertainty off a latent space. We read it off the gap between two inference modes of one model: with event context and without it. This gap is a per-slot analogue of the KL term that diagnoses posterior collapse. It needs no variational inference, no injective decoder, and no auxiliary loss regressor.
Research Program
The base model is TabEAE from Revisiting Event Argument Extraction: Can EAE Models Learn Better When Being Aware of Event Co-occurrences?. TabEAE fills a table whose rows are event triggers and whose columns are argument roles. A structure-aware attention mask decides whether each row can see the others. That paper already trains and tests under Single-Single, Multi-Multi and Multi-Single schemes, so one network can be queried with or without event context at no architectural cost. This is the property we need. The model also supplies the method we reuse.
The first workstream is a diagnostic. Let $p^{\text{multi}}_{e,r}$ be the model's distribution over candidate spans (plus "no argument") for role $r$ of event $e$ when all events in the document are visible. Let $p^{\text{single}}_{e,r}$ be the same distribution when only event $e$ is visible. The structural reliance of that slot is
$$ D_{e,r} = \mathrm{JS}\big(p^{\text{multi}}_{e,r} \,\|\, p^{\text{single}}_{e,r}\big), $$
where JS is the Jensen–Shannon divergence. JS is bounded and symmetric, so values are comparable across slots.
$D$ is not the error itself, because cross-event context often helps. The diagnostic question is whether high-$D$, high-confidence slots are wrong more often than low-$D$ slots of the same confidence. We also ask whether this effect grows as labels shrink and as roles get rarer. If the answer is yes, structural collapse is real and measurable. If high-$D$ slots are mostly corrections, where context fixes an evidence-only mistake, then the joint model is behaving well and the acquisition idea weakens. We will report that outcome as a negative result.
The second workstream is acquisition. Each document is scored by the mean of its top-$m$ slot divergences, weighted by the joint view's confidence:
$$ s(d) = \frac{1}{m}\sum_{(e,r)\in \text{top-}m(d)} D_{e,r}\cdot \max p^{\text{multi}}_{e,r}. $$
The top-$m$ pooling follows the finding in Active Learning for Event Extraction with Memory-based Loss Prediction Model that averaging over a whole sentence washes out the few slots that matter. The confidence weight targets the collapse signature: the model is sure, and its certainty comes from context.
We select whole documents and annotate every event in each. TabEAE found that co-occurring events regularize one another during training. Sentence-level greedy selection would break those chains, which is the concern raised as an open question in the file's third gap. We will test this directly against sentence-level selection, normalizing cost by the number of annotated arguments.
The third workstream borrows the standard fix for posterior collapse. Bowman et al. used word dropout, hiding decoder inputs so the model must use the latent code. Our analogue is context dropout. During training we randomly hide co-occurring event rows, so that each slot must sometimes be predicted from text alone. This keeps the single-event view competent, which makes $D$ meaningful. It may also reduce structural collapse by itself. That possibility is a real competitor to the acquisition function, and the ablations separate the two effects.
Baselines
The naive reference point is random document sampling with the same TabEAE backbone, the same Multi-Multi training, and the same per-round fine-tuning schedule. Every acquisition method spends the same budget of annotated arguments per round on the same unlabeled pools, which are carved from the training splits of ACE05, RAMS, WikiEvents and MLEE. The same backbone and hyperparameters are used throughout, with a single tuning pass done once on the random-sampling curve and then frozen for all methods. Five seeds are run per method. Beating random means a statistically separated area under the F1-versus-budget curve.
The strongest uncertainty competitor is MC dropout over TabEAE's slot distributions, using the same top-$m$ document aggregation. Wu et al. and Tsymbalov et al. report that MC dropout variance tracks model ignorance better than geometric diversity measures. That makes it the most credible cheap rival. We give it the same number of forward passes as our method's two inference modes, and also a generous 10-sample setting, so that a loss cannot be blamed on a compute mismatch. Plain predictive entropy over the Multi-Multi view runs alongside it.
The strongest existing event-specific active learner is MBLP from Active Learning for Event Extraction with Memory-based Loss Prediction Model. MBLP trains a loss-prediction head with a memory of past selections and ranks samples by their top-$m$ predicted token losses. That paper reports that on its Financial dataset MBLP reaches 90% of full-data performance with 11% of the data, against 13% for the next best active learning method and 22% for random sampling. We reimplement MBLP's loss head and memory module on top of TabEAE so the backbone is shared. Beating it means reaching 90% of full-data Arg-C F1 with a smaller labeled fraction on the English benchmarks.
There is no reported result found on Lacuna for any active learning method on ACE05, RAMS, WikiEvents or MLEE. Our curves therefore anchor to full-data numbers instead. EDGE reports 70.6 Arg-C F1 on ACE05 in Intra-Event and Inter-Event Dependency-Aware Graph Network for Event Argument Extraction. TabEAE reports a 2.7-point Arg-C gain on MLEE. Our own full-data TabEAE run defines the "90% of full performance" target.
Three ablations isolate the key ingredient. The first uses our score without context-dropout training. The second uses context-dropout training with MC-dropout acquisition, which tests whether the training fix alone explains the gains. The third uses the unweighted JS score, which tests whether the confidence weighting matters. We add a variant that filters out documents whose single-view entropy is extreme. It follows the finding in Privacy-preserving and Uncertainty-aware Federated Trajectory Prediction for Connected Autonomous Vehicles that maximum-uncertainty selection picks unlearnable noise. It tells us whether high-disagreement documents are informative or just messy.
Why not just…?
Why not just run MC dropout or entropy on TabEAE? This is the most serious threat, and it is cheaper in implementation effort. Our argument is mechanistic. A prediction driven by the co-occurrence prior is produced by weights that every dropout sample shares, so it should look certain under MC dropout even when it is wrong. The diagnostic workstream tests this directly. We will bin slots by MC-dropout variance and check whether high-$D$ slots within the low-variance bin carry elevated error. If they do not, MC dropout already captures the signal, and our acquisition curves will show no gain over it. In that case the honest conclusion is that structural collapse is not a practical problem for active learning.
Why not just use MBLP's learned loss predictor? MBLP regresses token losses. It has no notion of which part of a prediction comes from cross-event structure, and its "delayed training" means it learns from losses the base model has already incurred, on the distribution already labeled. Atypical event structures are rare by definition, so a loss regressor trained on typical ones has little reason to flag them. The atypical-structure test slice described below measures this. MBLP could still win on overall F1 through better batch diversity from its memory module. We therefore also report our score combined with MBLP's memory-based redundancy filter.
Why not just put a latent variable on the event structure and use its posterior variance, guarded by an identifiable decoder? This is the obvious combination of the three directions. It is less sound than it looks. An amortized posterior's variance describes how ambiguous the input is under the model. It does not describe how ignorant the model is. That is the aleatoric/epistemic conflation flagged in Peng et al.. Keeping the latent alive with the Brenier-map decoders of Posterior Collapse and Latent Variable Non-identifiability adds second-order backpropagation that the paper itself identifies as costly, and its evidence comes from density-estimation benchmarks rather than structured extraction. Image-domain latent active learners such as Variational Adversarial Active Learning and Task-Aware Variational Adversarial Active Learning avoid the autoregressive conditions under which collapse arises. We take the diagnosis from that literature and leave out the machinery.
Why not just drop dependency modeling and use a single-event extractor like PAIE? PAIE predicts each event's arguments from prompts and bipartite matching, with no cross-event interaction. It cannot collapse onto other events, but it also forgoes the co-occurrence gains documented by TabEAE and EDGE. We run a PAIE-backbone active learning curve with entropy acquisition. If it matches the joint model at low budgets, that would itself show that dependency modeling only pays off once enough labels exist.
Proposed Experiments
All experiments simulate pool-based active learning on ACE05, RAMS, WikiEvents and MLEE. Each starts from a seed set of 2% of training documents and runs ten acquisition rounds. Each round labels a fixed number of arguments' worth of documents, and the model is retrained from the pretrained BART checkpoint every round. MLEE is the primary testbed because its dense, nested events are where TabEAE found cross-event context matters most. If structural collapse exists, it should appear most strongly there. The Chinese Military, Legal and Financial sets from the MBLP paper are a secondary check if they can be obtained, since they would allow a like-for-like comparison with MBLP's reported 11% figure.
The metrics are Arg-I and Arg-C F1 against labeled budget, area under that curve, the labeled fraction needed to reach 90% of full-data Arg-C F1, and Arg-C F1 restricted to roles in the bottom frequency quartile. We also report confident-error rate: the share of slots with joint-view confidence above 0.9 that are wrong.
Two checks decide the outcome. The first is the existence test. Before any active learning, we train TabEAE at 5%, 20% and 100% of the data. We construct an atypical-structure slice of the test set, made of documents whose gold role pattern for some event differs from the majority co-occurrence pattern seen in training. On this slice we check whether high-$D$ confident slots have higher error than matched low-$D$ slots, and whether that gap shrinks with more data. If this fails, the acquisition claim has no foundation, and we stop and report.
The second is the attribution test. Structural-disagreement acquisition with context dropout must beat both MC dropout with context dropout and MBLP on area under the curve on MLEE and at least two of the other three benchmarks. The gain must be larger on rare roles and on the atypical slice than overall. If context dropout with plain MC dropout matches it, the contribution reduces to a training regularizer motivated by posterior collapse. That is a smaller result, but still a useful one, and we would present it as such.
References
- Generating Sentences from a Continuous Space
- Posterior Collapse and Latent Variable Non-identifiability
- Revisiting Event Argument Extraction: Can EAE Models Learn Better When Being Aware of Event Co-occurrences?
- Active Learning for Event Extraction with Memory-based Loss Prediction Model
- Uncertainty-Aware Balancing for Multilingual and Multi-Domain Neural Machine Translation Training
- Dropout-Based Active Learning for Regression
- Intra-Event and Inter-Event Dependency-Aware Graph Network for Event Argument Extraction
- Privacy-Preserving and Uncertainty-Aware Federated Trajectory Prediction for Connected Autonomous Vehicles
- Variational Adversarial Active Learning
- Task-Aware Variational Adversarial Active Learning
- Prompt for Extraction? PAIE: Prompting Argument Interaction for Event Argument Extraction