Argument-Centric Causal Intervention Method for Mitigating Bias in Cross-Document Event Coreference Resolution

Fuyuan Wei; Hongzhen Lv; Jiaren Peng; Liejun Wang; Long Yao; Wenzhong Yang; Xiaoming Tao; Yabo Yin

REVIEW 6 major objections 6 minor 53 references

Argument-Centric Causal Intervention Method for Mitigating Bias in Cross-Document Event Coreference Resolution

T0 review · 6 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that trigger-word matching acts as a confounder in cross-document event coreference, and that backdoor-adjusted counterfactual reasoning can subtract its bias, yielding state-of-the-art CoNLL F1 of 88.4% on ECB+ and…

desk verdict The ACCI debiasing idea is plausible and the case studies are nice, but the headline SOTA number is a test-set selection artifact and the causal language outruns what is actually computed. read the letter →

arxiv 2506.01488 v1 pith:4ZSITON7 submitted 2025-06-02 cs.CL cs.IR

Long Yao , Wenzhong Yang , Yabo Yin , Fuyuan Wei , Hongzhen Lv , Jiaren Peng , Liejun Wang , Xiaoming Tao This is my paper

classification cs.CLcs.IR

keywords EventCoreferenceResolutionCross-DocumentCausalInterventionBackdoorAdjustmentCounterfactualReasoningSpuriousCorrelationMitigationTriggerBiasArgument-AwareEnhancement

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Cross-document event coreference models are known to lean on the superficial similarity of trigger words (the verbs or nouns that signal an event), which misleads them when two mentions share a trigger but mean different things, or use different triggers for the same event. This paper claims that this failure is a causal one: trigger-word matching is a confounder that opens a backdoor path between the input pair and the predicted coreference label. The proposed ACCI framework applies backdoor adjustment with the do-operator, then builds a counterfactual, trigger-only version of each mention pair to estimate a bias score and an argument-only version to produce an argument-centric score; the final prediction adds the argument-centric score and subtracts the bias score. The paper reports state-of-the-art CoNLL F1 of 88.4% on ECB+ and 85.2% on GVC, with ablations showing that removing either component degrades both benchmarks.

What carries the argument

The load-bearing object is the trigger-matching confounder $T$ and its counterfactual proxy: the trigger-only embedding $h_{cf} = \mathrm{Cross\_Encoder}(x_{t1}, x_{t2} | X_{arg} = \varnothing)$, which is supposed to retain only the lexical features of the triggers. A linear projection of this embedding produces the bias score $S_{bias}$ that is subtracted at inference, while the argument-aware branch supplies the paired argument-only score $P_C$. The additive separability identity $y = P_F + \alpha P_C - \beta S_{bias}$ is the mechanism that carries the claim, because it asserts that the spurious signal and the argument signal are separable and that subtracting the former recovers the causal effect.

What would settle it

Construct a test set where every non-coreferent pair shares an identical trigger and every coreferent pair uses different triggers; if ACCI's debiased predictions do not separate the two classes better than the factual baseline, then the trigger-only embedding does not isolate the confound.

Watch

Extended reading notes

Core claim

ACCI's central claim is that the lexical trigger-matching signal $T$ acts as a confounder on the path $X \leftarrow T \rightarrow Y$ inside a pairwise event coreference classifier, and that the backdoor-adjusted interventional distribution $P(Y|do(X)) = \sum_t P(Y|X,A,T=t)P(T=t)$ isolates the argument-driven causal effect. To make this operational, the framework constructs a counterfactual, trigger-only embedding $h_{cf}$ (argument context masked out), derives from it a scalar bias score $S_{bias}$, and builds an argument-only prediction $P_C$ by replacing the trigger with an empty-trigger placeholder. The final debiased inference is $y = P_F + \alpha P_C - \beta S_{bias}$, with the counterfactual module deliberately excluded from backpropagation, and the training objective is $L = L_F + \alpha L_C$. On this basis the paper reports CoNLL F1 of 88.4% on ECB+ and 85.2% on GVC, with ablation studies showing that removing either the trigger-bias term or the argument-enhancement term lowers both scores.

Load-bearing premise

The entire debiasing procedure rests on the assumption that an embedding built only from trigger words, with all argument context masked out, isolates exactly the spurious bias, so that subtracting its score from the factual prediction removes the confound and nothing else.

Editorial extensions

If this is right

State-of-the-art CoNLL F1 on both benchmarks: 88.4% on ECB+ and 85.2% on GVC, outperforming the strongest reported baselines by 2.0% on ECB+ and staying within 0.6% of the best GVC baseline while topping B3 and LEA F1 there.
Causal debiasing can replace counterfactual data augmentation: ACCI achieves its gains without new labeled data or heuristic filtering, and without changing the underlying training procedure.
Trigger words carry both a spurious and a genuine signal: the beta sensitivity analysis shows optimal debiasing at $\beta \approx 0.2$–$0.3$, with full removal ($\beta = 1.0$) sharply degrading performance.
The argument-aware branch, trained on trigger-masked inputs, is sufficient to drive coreference decisions, implying that argument-level consistency is a usable decision signal on its own.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A testable extension: on a synthetic dataset where trigger-word overlap is anti-correlated with gold labels, ACCI should degrade less than its backbone; if it does not, the bias term is not capturing the confound.
The same additive debiasing recipe could transfer to any pairwise text task where a surface feature (e.g., token overlap) confounds the label, such as paraphrase detection or stance classification.
Because the counterfactual branch never receives gradients, its parameters are only shaped by the reconstruction of $h_{cf}$; a direct comparison against a version where the branch is also trained with a dedicated objective would reveal whether the bias estimate is being learned or merely memorized.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

You should know two things about this one. The architecture itself is not crazy: masking triggers, encoding the argument-only branch, and subtracting a trigger-only bias score is a reasonable debiasing recipe, and unlike Ding2024 it does not require extra data annotation. But the empirical headline that carries the paper — the 88.4 CoNLL F1 on ECB+ and the SOTA claim — does not survive contact with the experimental section.

What is genuinely new: the two-module design (counterfactual trigger-masking plus argument-aware enhancement) applied to CD-ECR, with a single shared encoder and no modification to the training loss except adding the auxiliary argument-only loss. The case studies in Fig. 10 and Fig. 11 are actually illustrative and show the model behaving sensibly. The comparison against two backbone architectures and the ablation are also more thorough than most papers in this area, and the code link is there.

Now the soft spots, in proportion. First, and most seriously, Section 5.6 sweeps the debiasing coefficient beta from 0.0 to 1.0 on the ECB+ test split and reports the maximum as the headline result. That is test-set selection, not a held-out evaluation. The baselines were not given this advantage, so the “state-of-the-art” comparison is not fair, and the central empirical claim collapses until this is re-done with a dev-set selection or a fixed beta.

Second, the causal machinery is mostly decorative. Equation (10) presents a backdoor adjustment, but the algorithm never computes P(Y|do(X)); the final prediction is a linear combination of three MLP outputs, with alpha and beta fit after training. Calling this “unbiased estimates of causal effects” is not supported by the math in the paper.

Third, the abstract claims SOTA on GVC with 85.2, but Table 4 shows Chen2025 at 85.8. That is an internal contradiction. Fourth, the ablation in Table 6 reports GVC numbers around 79.0, which correspond to the ACCI-Base row in Table 5, not the headline 85.2; the table should say which configuration it is ablating.

Who gets value from this: researchers working on event coreference who want a reproducible trigger-debiasing baseline, and anyone teaching why test-set hyperparameter selection invalidates a SOTA claim. This is not a desk-reject; it deserves a serious referee, but the referee should be instructed to verify the beta selection and the SOTA claims. The paper is salvageable: fix the selection procedure, correct the GVC SOTA statement, clarify the ablation setup, and soften the causal language.

Referee Report

6 major / 6 minor

Summary. The paper proposes ACCI (Argument-Centric Causal Intervention), a debiasing framework for cross-document event coreference resolution. The authors model trigger-word matching as a confounder in a structural causal graph, propose a backdoor adjustment in Eq. (10), and implement a counterfactual reasoning module that masks triggers to estimate a bias term and an argument-aware module that masks context to produce an argument-only prediction. The final prediction is a linear combination y = P_F + αP_C − βS_bias (Eq. 18). Experiments on ECB+ and GVC report CoNLL F1 of 88.4 and 85.2, which the abstract describes as state-of-the-art. The paper also provides case studies, hyperparameter analysis, and a GitHub link.

Significance. If the causal claims held, the paper would make a useful contribution: it applies causal intervention and counterfactual reasoning to CD-ECR, avoids data augmentation costs, and reports strong benchmark numbers. The paper is transparent enough to expose its own selection procedure (Fig. 8), which is a credit, and it provides code and detailed case analyses. However, the evidence as presented does not establish the central claims: the headline number is selected on the test set, the headline configuration appears to use an undefined oracle setting, the backdoor adjustment in Eq. (10) is never computed by the system, and the GVC result is not state-of-the-art by the paper's own Table 4. The core idea may be salvageable as a heuristic debiasing module, but the causal interpretation and the performance claims need substantial revision.

major comments (6)

[§5.6, Fig. 8, Eq. (18)] The headline ECB+ result is a test-set maximum, not a held-out result. Section 5.6 sweeps β from 0.0 to 1.0 in increments of 0.05 on the ECB+ test split and reports the best CoNLL F1 (88.4) and B3 F1 (86.9). No development-set selection is described for β, and no separate selection procedure is given for α in Eq. (18). The baselines in Table 2 were not given this test-set selection advantage, so the claimed state-of-the-art comparison is not a fair holdout comparison.
[Tables 2 and 3] The ACCI row in Table 2 is identical to the ACCI-Base row under 'Ahmed2023-LH_Oracle' in Table 3 (MUC 91.8, B3 86.9, CEAF 86.4, LEA 81.0, CoNLL 88.4), yet the table caption and Section 5.4.1 do not define what 'Oracle' means or state that the headline ACCI uses this setting. The baselines in Table 2 are not all reported under that same setting; for example, the reproduced Held2021 baseline in Table 3 is 84.6 CoNLL F1, while Table 2 lists 85.7. If the oracle setting provides gold candidate pairs or gold topics, comparing ACCI in that setting to non-oracle baselines is not an apples-to-apples comparison.
[§4.3.2, Eq. (10), §4.5, Eq. (18)] The backdoor adjustment in Eq. (10) is never computed by the model. Eq. (10) defines P(Y|do(X)) = Σ_t P(Y|X,A,T=t)P(T=t), but the actual inference rule in Eq. (18) is a weighted sum and subtraction of three outputs of the same network: y = P_F + αP_C − βS_bias. The summation over t ∈ τ and the marginal P(T=t) do not appear in the algorithm, and no derivation connects the causal estimand to Eq. (18). As written, the claims that ACCI 'applies the backdoor criterion' and 'provides unbiased estimates of causal effects' are not supported by the implemented procedure.
[Table 4, Abstract] The abstract's statement that ACCI 'achiev[es] state-of-the-art performance' is contradicted by Table 4 on GVC. ACCI obtains CoNLL F1 of 85.2, while Chen2025 in the same table obtains 85.8. At most, the paper can claim state-of-the-art on ECB+; the GVC claim as stated is inaccurate.
[Table 6 vs Tables 4 and 5] The GVC ablation row in Table 6 is inconsistent with the paper's main GVC results. Table 6 reports ACCI on GVC as MUC 91.0, B3 79.4, CEAFe 66.6, LEA 73.7, CoNLL 79.0, whereas Table 4 reports ACCI on GVC as MUC 92.6, B3 86.6, CEAFe 76.4, LEA 81.7, CoNLL 85.2. The Table 6 numbers match the ACCI-Base row under 'Ahmed2023-LH_Oracle' in Table 5, not the ACCI configuration used for the headline GVC result (which uses Held2021). The ablation is therefore not performed on the same system whose headline result is claimed, so the reported drops cannot be interpreted as contributions to the headline GVC performance.
[§4.4, Eqs. (11)–(15)] The counterfactual decomposition relies on unvalidated assumptions. Eq. (11) assumes that a trigger-only embedding with X_arg = ∅ isolates the spurious trigger signal, and Eqs. (12) and (15) assume that learnable placeholder vectors Φ_c and Φ_E faithfully simulate absent context and absent triggers. No experiment checks these assumptions, for example by testing whether trigger-only predictions track lexical similarity, whether the argument-only prediction is invariant to trigger replacement, or whether the placeholder vectors behave like true missing information. Without such validation, the debiasing procedure is a heuristic interpolation rather than a causally identified adjustment.

minor comments (6)

[§3.2.2] The heading 'Casual Intervention' should read 'Causal Intervention'.
[Throughout] There are numerous typos, including 'latent ariables' in §4.3.1, 'frameworkk' in §6, 'jurty' in Fig. 1, 'pepole' and 'erathquake' in Fig. 6, and 'CACI' instead of 'ACCI' in Figs. 7 and 10.
[§4.4.2] The vector Φ is introduced twice as a learnable placeholder, and the relation between Φ, Φ_c, and Φ_E is not made explicit; the notation should be unified.
[Algorithm 1] The cluster merging threshold τ_cluster appears in Algorithm 1 but its value is not reported in the implementation details in §5.2.
[§5.6, Fig. 8] The β sweep is only reported on ECB+. Since GVC is also a headline result, the paper should either provide a comparable sweep on GVC or explicitly state that β is fixed to the same value for GVC.
[Ethical Consideration vs Abstract] The abstract states that implementation and materials are available at the GitHub URL, but the Ethical Consideration section says materials 'will be uploaded as soon as the paper is accepted'; these statements should be reconciled.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported ECB+ SOTA is a test-set grid maximum for β, not a fixed-hyperparameter held-out result.

fitted input called prediction [Section 5.6 (Hyperparameter analysis), Fig. 8; reported in Section 5.4.1, Table 2]
"To quantitatively assess the impact of the debiasing coefficient β, we varied its value from 0.0 to 1.0 in increments of 0.05 and evaluated the resulting CoNLL F1 and B3 F1 on the ECB+ dataset... When β is set within the range of 0.2–0.3, the model can effectively suppress misleading correlations while retaining essential information from trigger words, achieving optimal performance on both evaluation metrics."

The ECB+ test split (topics 36–45) is the same split used for the headline results in Tables 2 and 3; no development-set selection of β is described. The reported 88.4 CoNLL F1 / 86.9 B3 F1 are the peak values of a grid search over β in the final prediction equation (Eq. 18: y = P_F + αP_C − βS_bias) evaluated directly on the test set. The 'state-of-the-art' result is therefore the maximum of the test grid, not a held-out prediction with hyperparameters fixed before observing test labels, and the baselines in Table 2 were not given this test-set tuning advantage.

full rationale

The only concrete circular reduction I can exhibit in the paper is evaluative: the headline ECB+ result (88.4 CoNLL F1) is produced by sweeping the debiasing coefficient β on the ECB+ test split and reporting the optimum as ACCI's performance. That is a fitted-input-called-prediction pattern and it is load-bearing because the paper's central empirical claim is the SOTA number. The broader causal framing is also underderived—Eq. 10's backdoor adjustment is never instantiated in the algorithm, and Eq. 18 is a weighted linear combination of three outputs of the same shared-parameter model (P_F, P_C, S_bias) with S_bias defined as the spurious-bias score—but that is better described as an unsupported causal claim than as circularity, since Eq. 18 is not shown to be equivalent to Eq. 10 by construction. No load-bearing self-citation or imported uniqueness theorem is present. Because the β sweep directly selects the metric being reported as SOTA, a score of 6 is appropriate; absent that sweep, the causal gap would be a correctness/validity concern rather than a circularity finding.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central causal claims rest on an assumed graph, an unvalidated counterfactual construction, and two tuned coefficients. The free parameters alpha and beta are not fixed by theory and are selected based on test-set performance. The placeholder vectors are new learnable components introduced to make the masking procedure work.

free parameters (4)

alpha (loss and inference weight) = not reported
Balances factual and argument-only losses (Eq. 16) and combines predictions at inference (Eq. 18); tuned on performance, value not given in the paper.
beta (debiasing coefficient) = 0.2 to 0.3 (optimal range), swept 0.0 to 1.0
Controls subtraction of trigger-bias term (Eq. 18); tuned on the ECB+ test set, as shown in Fig. 8, rather than on a validation set.
Learnable placeholder vector Phi (empty trigger) = learned during training
Replaces trigger embeddings in the argument-only module (Eq. 15); introduced ad hoc to simulate the absence of trigger information.
Empty-context placeholder Phi_c = zero or learned placeholder
Used in counterfactual embedding construction (Eq. 12); the paper does not specify whether it is zero or learnable, and its behavior as an 'empty context' is unvalidated.

assumptions (4)

domain assumption T (trigger matching) is a confounder inducing path X <- T -> Y in the causal graph (Fig. 3b)
The entire causal framing rests on this graph; no empirical test shows that T plays this causal role.
ad hoc to paper Masking triggers or context yields valid counterfactual inputs that the cross-encoder interprets as the absence of that information
The counterfactual construction (Eqs. 11, 14, 15) assumes the model's representation of masked inputs corresponds to Pearl's counterfactual quantities, but no evidence supports this equivalence.
domain assumption A (argument semantics) is a mediator on the path X -> A -> Y rather than a confounder
In Eq. (10), A is included in the backdoor adjustment despite being a descendant of X, which is invalid under the standard backdoor criterion; this structural choice is asserted without justification.
domain assumption Standard datasets (ECB+, GVC) and splits are unbiased and comparable across baselines
Evaluation follows prior work, but the paper does not verify whether all compared methods use the same candidate retrieval setting, especially the Oracle topic setup used for the headline ECB+ result.

invented entities (2)

Empty-context placeholder Phi_c
purpose: Represents zero contextual input in counterfactual embedding (Eq. 12)
A new vector introduced to simulate absent context; there is no external validation that the model treats it as a true empty context.
Empty-trigger placeholder Phi_E
purpose: Replaces trigger embeddings to simulate argument-only inference (Eq. 15)
Learnable vector; the model's ability to treat it as a generic empty event is not validated outside the paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Argument-Centric Causal Intervention Method for Mitigating Bias in Cross-Document Event Coreference Resolution." pith.science (2026). https://pith.science/paper/4ZSITON7

@misc{pith2026250601488,
  author       = {Pith},
  title        = {Pith review of: Argument-Centric Causal Intervention Method for Mitigating Bias in Cross-Document Event Coreference Resolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZSITON7}},
  note         = {Machine review of arXiv:2506.01488}
}

read the original abstract

Cross-document Event Coreference Resolution (CD-ECR) is a fundamental task in natural language processing (NLP) that seeks to determine whether event mentions across multiple documents refer to the same real-world occurrence. However, current CD-ECR approaches predominantly rely on trigger features within input mention pairs, which induce spurious correlations between surface-level lexical features and coreference relationships, impairing the overall performance of the models. To address this issue, we propose a novel cross-document event coreference resolution method based on Argument-Centric Causal Intervention (ACCI). Specifically, we construct a structural causal graph to uncover confounding dependencies between lexical triggers and coreference labels, and introduce backdoor-adjusted interventions to isolate the true causal effect of argument semantics. To further mitigate spurious correlations, ACCI integrates a counterfactual reasoning module that quantifies the causal influence of trigger word perturbations, and an argument-aware enhancement module to promote greater sensitivity to semantically grounded information. In contrast to prior methods that depend on costly data augmentation or heuristic-based filtering, ACCI enables effective debiasing in a unified end-to-end framework without altering the underlying training procedure. Extensive experiments demonstrate that ACCI achieves CoNLL F1 of 88.4% on ECB+ and 85.2% on GVC, achieving state-of-the-art performance. The implementation and materials are available at https://github.com/era211/ACCI.

Figures

Figures reproduced from arXiv: 2506.01488 by the authors.

**Figure 3.** Structural Causal Graph for Event Coreference Resolution. This figure illustrates causal relationships in ECR using a Structural Causal Model. Key variables include event mention pairs (𝑋), coreference labels (𝑌 ), and trigger word matching (𝑇 ) as a confounder. Path 𝑋 ← 𝑇 → 𝑌 shows how trigger words simultaneously influence mention representation and coreference judgment, creating spurious correlations. Subfigures… view at source ↗

**Figure 1.** Examples of Event Mentions in Within-Document and Cross-Document Scenarios under the ECR Task. Document 1 contains mentions 𝑒1 and 𝑒2 , while document 2 includes mentions 𝑒3 and 𝑒4 . In WD-ECR, the model evaluates intra-document pairs like (𝑒3 , 𝑒4 ). In CDECR, it analyzes crossdocument pairs like (𝑒1 , 𝑒4 ) to cluster coreferential events across texts. C-ECB+ NC-ECB+ C-GVC NC-GVC 0 20 40 60 80 100 Percentage(%)… view at source ↗

**Figure 2.** Distribution of “trigger word matching” between mention pairs in the ECB+ and GVC training sets (The left is ECB+ and the right is GVC), with 𝐶 and 𝑁𝐶 denoting coreferential and non-coreferential pairs, respectively. 𝐿𝑒𝑥𝑆𝑖𝑚 versus 𝐿𝑒𝑥𝐷𝑖𝑣 indicating lexically similar or dissimilar trigger words. underscored that current models tend to overly depend on the lexical matching of trigger words, and this characteristic is … view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Illustrative of a causal graph, where 𝑋, 𝑌 , and 𝐶 represent the cause, outcome, and mediator, respectively. and a set of edges 𝜉 that represent the causal links between these variables. Within a causal graph, several common causal structures are frequently observed, i…

**Figure 6.** Figure 6: The overall architecture of the ACCI framework. 4.1. Task Definition Given a document set  = {𝑑1 , 𝑑2 , …, 𝑑𝑚}, where 𝑚 = ||. Each document 𝑑𝑖 contains multiple event mentions. The set of all event mentions in  constitutes the event mention set 𝐸 = {𝑒1 , 𝑒2 , …, 𝑒𝑛…

**Figure 7.** Figure 7: Distribution of correctly resolved mentions (True Positives) across cluster-size thresholds on the ECB+ and GVC datasets, comparing the ACCI model with baseline. lexical cues. In contrast, baseline models tend to rely heavily on trigger word matching and superficial le…

**Figure 9.** Figure 9: Training loss curves of BCELoss and BCEWithLogitsLoss under the ACCI framework. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

**Figure 10.** Figure 10: Three cross-document mention-pair cases from the ECB+ dataset, with event triggers marked between < 𝑚 > and < ∕𝑚 >, “Non-Coreference” denoting non-coreferential pairs and “Coreference” denoting coreferential pairs; A bar chart is included for Case 1 to illustrate the …

**Figure 11.** Figure 11: Case studies illustrating the impact of lexical trigger bias and argument sensitivity. ACCI shows improved robustness over the baseline by correctly handling semantically similar triggers, reducing sensitivity to superficial trigger matches, and detecting changes in e…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 49 canonical work pages

[1]

Semeval-2018 task 5: Counting events and participants in the long tail

Marten Postma, Filip Ilievski, and Piek Vossen. Semeval-2018 task 5: Counting events and participants in the long tail. InProceedings of The 12th International Workshop on Semantic Evaluation, pages 70–80, 2018

work page 2018
[2]

Event coreference for information extraction

Kevin Humphreys, Robert Gaizauskas, and Saliha Azzam. Event coreference for information extraction. InOperational Factors in Practical, Robust Anaphora Resolution for Unrestricted Texts, 1997

work page 1997
[3]

Hot- potqa: A dataset for diverse, explainable multi-hop question answer- ing.arXiv preprint arXiv:1809.09600, 2018

Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hot- potqa: A dataset for diverse, explainable multi-hop question answer- ing.arXiv preprint arXiv:1809.09600, 2018

arXiv 2018
[4]

InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4179–4189, 2019

Shany Barhom, Vered Shwartz, Alon Eirew, Michael Bugert, Nils Reimers,andIdoDagan.Revisitingjointmodelingofcross-document entity and event coreference resolution. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4179–4189, 2019

work page 2019
[5]

2* n is better than n2: Decomposing event coreferenceresolutionintotwotractableproblems

Shafiuddin Rehan Ahmed, Abhijnan Nath, James H Martin, and Nikhil Krishnaswamy. 2* n is better than n2: Decomposing event coreferenceresolutionintotwotractableproblems. InFindingsofthe Association for Computational Linguistics: ACL 2023, pages 1569– 1583, 2023

work page 2023
[6]

Pairwise representa- tion learning for event coreference

Xiaodong Yu, Wenpeng Yin, and Dan Roth. Pairwise representa- tion learning for event coreference. In Vivi Nastase, Ellie Pavlick, Mohammad Taher Pilehvar, José Camacho-Collados, and Alessan- dro Raganato, editors,Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, *SEM@NAACL-HLT 2022, Seattle, WA, USA, July 14-15, 2022, pages ...

work page 2022
[7]

A rationale-centric counterfactual data augmen- tation method for cross-document event coreference resolution

Bowen Ding, Qingkai Min, Shengkun Ma, Yingjie Li, Linyi Yang, and Yue Zhang. A rationale-centric counterfactual data augmen- tation method for cross-document event coreference resolution. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)...

work page 2024
[8]

What happens before and after: Multi-event commonsense in event coreference resolution

Sahithya Ravi, Chris Tanner, Raymond Ng, and Vered Shwartz. What happens before and after: Multi-event commonsense in event coreference resolution. InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1708–1724, 2023

work page 2023

Show all 53 references

[9]

John Wiley & Sons, 2016

Judea Pearl, Madelyn Glymour, and Nicholas P Jewell.Causal inference in statistics: A primer. John Wiley & Sons, 2016

2016
[10]

Cross-document event coreference: Annotations, experiments, and observations

Amit Bagga and Breck Baldwin. Cross-document event coreference: Annotations, experiments, and observations. InCoreference and Its Applications, 1999

1999
[11]

Joint learning for event coreference resolu- tion

Jing Lu and Vincent Ng. Joint learning for event coreference resolu- tion. InProceedingsofthe55thAnnualMeetingoftheAssociationfor Computational Linguistics (Volume 1: Long Papers), pages 90–101, 2017

2017
[12]

Employing multiple decomposable attention networks to resolve event coreference

Jie Fang, Peifeng Li, and Guodong Zhou. Employing multiple decomposable attention networks to resolve event coreference. In CCF International Conference on Natural Language Processing and Chinese Computing, pages 246–256. Springer, 2018

2018
[13]

Corefprompt: Prompt- based event coreference resolution by measuring event type and argument compatibilities

Sheng Xu, Peifeng Li, and Qiaoming Zhu. Corefprompt: Prompt- based event coreference resolution by measuring event type and argument compatibilities. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 15440– 15452, 2023

2023
[14]

Learning event-aware measures for event coreference resolution

Yao Yao, Zuchao Li, and Hai Zhao. Learning event-aware measures for event coreference resolution. InFindings of the Association for Computational Linguistics: ACL 2023, pages 13542–13556, 2023

2023
[15]

Avi Caciularu, Arman Cohan, Iz Beltagy, Matthew E Peters, Arie Cattan,andIdoDagan.Cdlm:Cross-documentlanguagemodeling.In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2648–2662, 2021

2021
[16]

Cross-document coreference resolution over predicted men- tions

Arie Cattan, Alon Eirew, Gabriel Stanovsky, Mandar Joshi, and Ido Dagan. Cross-document coreference resolution over predicted men- tions. InFindings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 5100–5107, 2021

2021
[17]

End- to-end neural event coreference resolution.Artificial Intelligence, 303:103632, 2022

YaojieLu,HongyuLin,JialongTang,XianpeiHan,andLeSun. End- to-end neural event coreference resolution.Artificial Intelligence, 303:103632, 2022. Yao et al.:Preprint submitted to ElsevierPage 18 of 20 Argument-Centric Causal Intervention Method for Mitigating Bias in Cross-Docume...

2022
[18]

Focus on what matters: Applying discourse coherence theory to cross document coreference

William Held, Dan Iter, and Dan Jurafsky. Focus on what matters: Applying discourse coherence theory to cross document coreference. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1406–1417, 2021

2021
[19]

Cross- document event coreference resolution on discourse structure

Xinyu Chen, Sheng Xu, Peifeng Li, and Qiaoming Zhu. Cross- document event coreference resolution on discourse structure. In Proceedingsofthe2023ConferenceonEmpiricalMethodsinNatural Language Processing, pages 4833–4843, 2023

2023
[20]

Enhancing cross-document event coreference resolution by discourse structure and semantic informa- tion

Qiang Gao, Bobo Li, Zixiang Meng, Yunlong Li, Jun Zhou, Fei Li, Chong Teng, and Donghong Ji. Enhancing cross-document event coreference resolution by discourse structure and semantic informa- tion. InProceedings of the 2024 Joint International Conference on Computational Lingu...

2024
[21]

Improving cross- document event coreference resolution by discourse coherence and structure.Information Processing & Management, 62(4):104085, 2025

Xinyu Chen, Peifeng Li, and Qiaoming Zhu. Improving cross- document event coreference resolution by discourse coherence and structure.Information Processing & Management, 62(4):104085, 2025

2025
[22]

Hypergraph convolutional networks with multi-ordering relations for cross-document event coreference resolution.Information Fusion, 115:102769, 2025

Wenbin Zhao, Yuhang Zhang, Di Wu, Feng Wu, and Neha Jain. Hypergraph convolutional networks with multi-ordering relations for cross-document event coreference resolution.Information Fusion, 115:102769, 2025

2025
[23]

Linear cross-document event coreference resolution with x-amr

Shafiuddin Rehan Ahmed, George Arthur Baker, Evi Judge, Michael Reagan,KristinWright-Bettner,MarthaPalmer,andJamesHMartin. Linear cross-document event coreference resolution with x-amr. In Proceedings of the 2024 Joint International Conference on Com- putational Linguistics, L...

2024
[24]

Okay, let’s do this! modeling event coreference with generated rationales and knowledge distillation

Abhijnan Nath, Shadi Manafi Avari, Avyakta Chelle, and Nikhil Krishnaswamy. Okay, let’s do this! modeling event coreference with generated rationales and knowledge distillation. InProceedings of the 2024 Conference of the North American Chapter of the Associa- tion for Computa...

2024
[25]

Causality:models,reasoning,andinference, by judea pearl, cambridge university press, 2000.Econometric Theory, 19(4):675–685, 2003

LelandGersonNeuberg. Causality:models,reasoning,andinference, by judea pearl, cambridge university press, 2000.Econometric Theory, 19(4):675–685, 2003

2000
[26]

A causaldebiasingframeworkforunsupervisedsalientobjectdetection

Xiangru Lin, Ziyi Wu, Guanqi Chen, Guanbin Li, and Yizhou Yu. A causaldebiasingframeworkforunsupervisedsalientobjectdetection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 1610–1619, 2022

2022
[27]

Causal attention for vision-language tasks

Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847– 9857, 2021

2021
[28]

Counterfactualinferencefortextclassificationdebiasing

Chen Qian, Fuli Feng, Lijie Wen, Chunping Ma, and Pengjun Xie. Counterfactualinferencefortextclassificationdebiasing. InProceed- ingsofthe59thAnnualMeetingoftheAssociationforComputational Linguistics and the 11th International Joint Conference on Natural Language Processing (V...

2021
[29]

Unbiased scene graph generation from biased training

Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716–3725, 2020

2020
[30]

Causal inference in statistics: An overview

Judea Pearl. Causal inference in statistics: An overview. 2009

2009
[31]

Cambridge university press, 2009

Judea Pearl.Causality. Cambridge university press, 2009

2009
[32]

Causal intervention for leveraging popularity bias in recommendation

Yang Zhang, Fuli Feng, Xiangnan He, Tianxin Wei, Chonggang Song, Guohui Ling, and Yongdong Zhang. Causal intervention for leveraging popularity bias in recommendation. InProceedings of the 44th international ACM SIGIR conference on research and development in information retri...

2021
[33]

Clicks can be cheating: Counterfactual recommendation formitigatingclickbaitissue

Wenjie Wang, Fuli Feng, Xiangnan He, Hanwang Zhang, and Tat- Seng Chua. Clicks can be cheating: Counterfactual recommendation formitigatingclickbaitissue. InProceedingsofthe44thInternational ACMSIGIRConferenceonResearchandDevelopmentinInformation Retrieval, pages 1288–1297, 2021

2021
[34]

De-biasing distantlysupervisednamedentityrecognitionviacausalintervention

Wenkai Zhang, Hongyu Lin, Xianpei Han, and Le Sun. De-biasing distantlysupervisednamedentityrecognitionviacausalintervention. InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Confer- ence on Natural Lang...

2021
[35]

Counterfactually-augmented snli training data does not yield better generalization than unaugmented data

William Huang, Haokun Liu, and Samuel Bowman. Counterfactually-augmented snli training data does not yield better generalization than unaugmented data. InProceedings of the First Workshop on Insights from Negative Results in NLP, pages 82–87, 2020

2020
[36]

Deconfounded image captioning: A causal retrospect.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12996–13010, 2021

Xu Yang, Hanwang Zhang, and Jianfei Cai. Deconfounded image captioning: A causal retrospect.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12996–13010, 2021

2021
[37]

Context de-confounded emotion recognition

Dingkang Yang, Zhaoyu Chen, Yuzheng Wang, Shunli Wang, Mingcheng Li, Siao Liu, Xiao Zhao, Shuai Huang, Zhiyan Dong, Peng Zhai, et al. Context de-confounded emotion recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19005–19...

2023
[38]

What if the tv was off? examining counterfactual reasoning abilities of multi-modal language models

Letian Zhang, Xiaotong Zhai, Zhongkai Zhao, Yongshuo Zong, Xin Wen, and Bingchen Zhao. What if the tv was off? examining counterfactual reasoning abilities of multi-modal language models. InProceedingsoftheIEEE/CVFConferenceonComputerVisionand Pattern Recognition, pages 21853–...

2024
[39]

Causal intervention and counterfactual reasoning for multi-modal fakenewsdetection

Ziwei Chen, Linmei Hu, Weixin Li, Yingxia Shao, and Liqiang Nie. Causal intervention and counterfactual reasoning for multi-modal fakenewsdetection. InProceedingsofthe61stAnnualMeetingofthe Association for Computational Linguistics (Volume 1: Long Papers), pages 627–638, 2023

2023
[40]

Enhancing event causality identification with counterfactual reasoning

Feiteng Mu and Wenjie Li. Enhancing event causality identification with counterfactual reasoning. InProceedings of the 61st Annual MeetingoftheAssociationforComputationalLinguistics(Volume2: Short Papers), pages 967–975, 2023

2023
[41]

Bayes’ theorem

James Joyce. Bayes’ theorem. 2003

2003
[42]

Probabilities of causation: three counterfactual interpre- tationsandtheiridentification

Judea Pearl. Probabilities of causation: three counterfactual interpre- tationsandtheiridentification. InProbabilisticandcausalinference: the works of Judea Pearl, pages 317–372. 2022

2022
[43]

bag of events

Agata Cybulska and Piek Vossen. " bag of events" approach to event coreference resolution. supervised classification of event templates. Int. J. Comput. Linguistics Appl., 6(2):11–27, 2015

2015
[44]

Agata Cybulska and Piek Vossen. Using a sledgehammer to crack a nut?lexicaldiversityandeventcoreferenceresolution.InProceedings of the Ninth International Conference on Language Resources and Evaluation (LREC’14), pages 4545–4552, 2014

2014
[45]

Don’t annotate,butvalidate:Adata-to-textmethodforcapturingeventdata

PiekVossen,FilipIlievski,MartenPostma,andRoxaneSegers. Don’t annotate,butvalidate:Adata-to-textmethodforcapturingeventdata. InProceedingsoftheEleventhInternationalConferenceonLanguage Resources and Evaluation (LREC 2018), 2018

2018
[46]

Generalizing cross-documenteventcoreferenceresolutionacrossmultiplecorpora

Michael Bugert, Nils Reimers, and Iryna Gurevych. Generalizing cross-documenteventcoreferenceresolutionacrossmultiplecorpora. Computational Linguistics, 47(3):575–614, 2021

2021
[47]

Entity-basedcross-documentcoref- erencing using the vector space model

AmitBaggaandBreckBaldwin. Entity-basedcross-documentcoref- erencing using the vector space model. InCOLING 1998 Volume 1: The17thinternationalconferenceoncomputationallinguistics,1998

1998
[48]

Which coreference evaluationmetricdoyoutrust?aproposalforalink-basedentityaware metric

Nafise Sadat Moosavi and Michael Strube. Which coreference evaluationmetricdoyoutrust?aproposalforalink-basedentityaware metric. InProceedings of the 54th annual meeting of the association for computational linguistics, volume 1, pages 632–642. Association for Computational Li...

2016
[49]

A model-theoretic coreference scoring scheme

Marc Vilain, John D Burger, John Aberdeen, Dennis Connolly, and Lynette Hirschman. A model-theoretic coreference scoring scheme. InSixth Message Understanding Conference (MUC-6): Proceedings of a Conference Held in Columbia, Maryland, November 6-8, 1995, 1995

1995
[50]

On coreference resolution performance metrics

Xiaoqiang Luo. On coreference resolution performance metrics. In Proceedings of human language technology conference and confer- enceonempiricalmethodsinnaturallanguageprocessing,pages25– 32, 2005. Yao et al.:Preprint submitted to ElsevierPage 19 of 20 Argument-Centric Causal ...

2005
[51]

Pytorch.ProgrammingwithTensorFlow:solutionforedgecomputing applications, pages 87–104, 2021

SagarImambi,KollaBhanuPrakash,andGRKanagachidambaresan. Pytorch.ProgrammingwithTensorFlow:solutionforedgecomputing applications, pages 87–104, 2021

2021
[52]

Decoupled weight decay regular- ization.In7thInternationalConferenceonLearningRepresentations, ICLR2019,NewOrleans,LA,USA,May6-9,2019.OpenReview.net, 2019

Ilya Loshchilov and Frank Hutter. Decoupled weight decay regular- ization.In7thInternationalConferenceonLearningRepresentations, ICLR2019,NewOrleans,LA,USA,May6-9,2019.OpenReview.net, 2019

2019
[53]

Roberta: A robustly optimized bert pretraining approach

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, DanqiChen,OmerLevy,MikeLewis,LukeZettlemoyer,andVeselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. Yao et al.:Preprint submitted to ElsevierPage 20 of 20

1907 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Semeval-2018 task 5: Counting events and participants in the long tail

Marten Postma, Filip Ilievski, and Piek Vossen. Semeval-2018 task 5: Counting events and participants in the long tail. InProceedings of The 12th International Workshop on Semantic Evaluation, pages 70–80, 2018

work page 2018

[2] [2]

Event coreference for information extraction

Kevin Humphreys, Robert Gaizauskas, and Saliha Azzam. Event coreference for information extraction. InOperational Factors in Practical, Robust Anaphora Resolution for Unrestricted Texts, 1997

work page 1997

[3] [3]

Hot- potqa: A dataset for diverse, explainable multi-hop question answer- ing.arXiv preprint arXiv:1809.09600, 2018

Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and Christopher D Manning. Hot- potqa: A dataset for diverse, explainable multi-hop question answer- ing.arXiv preprint arXiv:1809.09600, 2018

arXiv 2018

[4] [4]

InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4179–4189, 2019

Shany Barhom, Vered Shwartz, Alon Eirew, Michael Bugert, Nils Reimers,andIdoDagan.Revisitingjointmodelingofcross-document entity and event coreference resolution. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4179–4189, 2019

work page 2019

[5] [5]

2* n is better than n2: Decomposing event coreferenceresolutionintotwotractableproblems

Shafiuddin Rehan Ahmed, Abhijnan Nath, James H Martin, and Nikhil Krishnaswamy. 2* n is better than n2: Decomposing event coreferenceresolutionintotwotractableproblems. InFindingsofthe Association for Computational Linguistics: ACL 2023, pages 1569– 1583, 2023

work page 2023

[6] [6]

Pairwise representa- tion learning for event coreference

Xiaodong Yu, Wenpeng Yin, and Dan Roth. Pairwise representa- tion learning for event coreference. In Vivi Nastase, Ellie Pavlick, Mohammad Taher Pilehvar, José Camacho-Collados, and Alessan- dro Raganato, editors,Proceedings of the 11th Joint Conference on Lexical and Computational Semantics, *SEM@NAACL-HLT 2022, Seattle, WA, USA, July 14-15, 2022, pages ...

work page 2022

[7] [7]

A rationale-centric counterfactual data augmen- tation method for cross-document event coreference resolution

Bowen Ding, Qingkai Min, Shengkun Ma, Yingjie Li, Linyi Yang, and Yue Zhang. A rationale-centric counterfactual data augmen- tation method for cross-document event coreference resolution. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)...

work page 2024

[8] [8]

What happens before and after: Multi-event commonsense in event coreference resolution

Sahithya Ravi, Chris Tanner, Raymond Ng, and Vered Shwartz. What happens before and after: Multi-event commonsense in event coreference resolution. InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1708–1724, 2023

work page 2023