Pith. sign in

REVIEW 5 major objections 3 minor 49 references

This paper claims that current large language models, while strong at detecting events, demonstrably fail at ordering event end times and constructing cross-document causal graphs, and that this is revealed by a new multi-granularity benchm

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

MiGUE-Bench is a 3,290-instance benchmark spanning event detection, relation reasoning, structure induction, and future prediction, showing LLMs are weakest at causal graph construction and end-time ordering.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Useful new benchmark, but the headline causal-graph deficit likely reflects an unvalidated propagated-gold-label construction more than a true capability boundary. the 5 major comments →

arxiv 2607.27654 v1 pith:B6MOSBLU submitted 2026-07-30 cs.CL cs.AI

From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language Models

Tao Wen , Shuai Shao , Pei Ke , Xu Han , Jie Zou , Guannan Li , Tao Tian , Jinjie Qiu
show 2 more authors
Lan Wang Ke Qin
This is my paper
classification cs.CL cs.AI
keywords event analysislarge language modelsbenchmarkevent detectionevent relation reasoningtemporal orderingcausal graph constructionretrieval-augmented generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that existing event benchmarks are too narrow—they test one document or one relation type at a time—and introduces MiGUE-Bench to measure the full arc of event analysis in one place: detecting event triggers, reasoning about temporal/causal/coreference/subevent relations within and across documents, inducing global event structures, and predicting future events. To build the benchmark at scale, the authors propose MiGUE-Pipeline, an LLM-driven annotation pipeline with self-correction, multi-model voting, and coreference-based relation propagation, yielding 3,290 instances from news sources. Experiments across twelve LLMs and two retrieval-augmented generation frameworks show that current models are strongest at event detection and coreference/subevent reasoning, but much weaker at ordering event end times and at constructing causal graphs, where accuracy approaches chance. The paper argues these results delineate the capability boundary of LLMs for event analysis and identify targets for future improvement.

Core claim

MiGUE-Bench is designed to span event analysis from single-document detection to cross-document prediction, organized around four tasks: MiGUE-Detection (multi-choice trigger selection with adversarial distractors), MiGUE-Reasoning (temporal, causal, coreference, and subevent relations, each in intra- and cross-document settings), MiGUE-Induction (interval-based start/end temporal ordering and causal-graph path selection), and MiGUE-Prediction (choosing the next event in a timestamped document sequence). The experiments show that while frontier closed- and open-source LLMs reach 0.84–0.86 Micro-F1 on detection and 0.96–0.97 on coreference reasoning, end-time ordering accuracy drops to roughl

What carries the argument

The load-bearing machinery is MiGUE-Pipeline, a four-stage LLM-driven annotation framework: document filtering via an LLM judge; event annotation via candidate trigger generation plus retrieval-augmented self-correction against six known failure types; relation annotation via constraint-aware multi-step reasoning with multi-model majority voting and escalation to a stronger model, plus propagate-via-coreference to densify cross-document relations; and cluster generation via a document-affinity graph partitioned with a community-detection algorithm. On top of this, the benchmark's task design carries the argument: decomposing temporal ordering into separate start/end permutations and framing

Load-bearing premise

The benchmark's ground-truth answers are generated by an automatic LLM pipeline with only a small manually checked sample; if these labels carry systematic biases—especially biases shared by the evaluated model families—the measured capability gaps could partly reflect annotator-model agreement rather than genuine event-analysis skill.

What would settle it

Take a random sample of MiGUE-Induction and MiGUE-Prediction instances and have independent human annotators label them; if human-machine agreement is no higher than the LLMs' near-chance scores on causal graphs, the ground truth is the weak link rather than the models. Alternatively, re-run the benchmark with option order and distractors reversed; a big accuracy drop would indicate shallow pattern matching rather than event understanding.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Current LLMs' event-analysis ability is lopsided: detection, coreference, and subevent reasoning are comparatively strong, while duration-aware temporal ordering and causal graph induction are far weaker.
  • RAG is not a reliable cure-all: it improves strong models on most tasks but can hurt weaker models by adding noise.
  • Model scale changes the benefit of long context: larger models gain steadily from longer instruction inputs, while smaller models peak and then decline.
  • The automatic annotation pipeline can produce event benchmarks at scale, reducing dependence on manual labeling, with spot-check precision/accuracy values reported on small samples.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the benchmark's labels are produced by LLMs, the reported accuracy gaps could partly measure how far the evaluated models are from the annotators' event ontology; a human-labeled subset would let readers separate model capability from label bias.
  • The end-time ordering gap suggests event duration is not being modeled; one testable extension is to supervise models on explicit interval predictions and see whether cross-document temporal reasoning improves.
  • If causal graph construction remains at chance even with strong retrieval, the bottleneck may be in how causal relations are defined (textual logical cues) rather than in retrieval; reformulating causality as a probabilistic schema-level relation would give a sharper test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 3 minor

Summary. The paper introduces MiGUE-Bench, a benchmark for multi-granularity event analysis, and MiGUE-Pipeline, an LLM-driven annotation pipeline. The benchmark comprises four tasks: event detection, relation reasoning, structure induction, and future prediction, at both single-document and cross-document levels. The authors evaluate a range of closed- and open-source LLMs and RAG methods, reporting that even strong models struggle with end-time ordering and causal graph construction, with the latter near chance. The paper also presents analyses of retrieval count and input length.

Significance. If the benchmark labels are reliable, MiGUE-Bench would fill a clear gap in event analysis evaluation by covering the full spectrum from atomic extraction to cross-document structure induction and prediction. The release of code and data is a strength, and the task taxonomy is well motivated. However, the validity of the capability measurements is not yet established given the small manual validation sets and the unsupported claim of full manual review.

major comments (5)
  1. [§3.3, §4.3.2, Table 4] The gold causal graphs are built by propagate-via-coreference transitivity rules (§3.3) with no validation of propagated edges. Since §4.3.2 defines correct options as maximal paths in the resulting DAG, any spurious edge changes the answer key. The near-chance scores on Causal Graph (e.g., Gemini-3-Pro 0.2159, Qwen3-30B 0.0530 in Table 4) may thus reflect an unsound gold standard rather than deficient LLM causal reasoning. Please validate a sample of the final DAGs and report agreement on propagated edges, or temper the 'critical deficiency' claim.
  2. [§4.5 vs §3.2, §3.3] §4.5 claims 'manual review on all the instances', but the only evidence is 200-event precision 0.87 (§3.2) and 100-relation accuracy 0.82 (§3.3). No annotator count, inter-annotator agreement, or audit trail is provided. For a benchmark intended to measure LLM capabilities, this is insufficient. Provide a detailed quality-control report or revise the claim.
  3. [§3.2, §3.3, Table 4] The same models that generate labels are evaluated: DeepSeek-V3.2 for triggers (§3.2); GPT-4o, DeepSeek-V3.2, Qwen3-Max with Claude-4.5-Opus arbitration for relations (§3.3). Table 4 evaluates these models. Scores may partly measure agreement with the annotation pipeline rather than general ability. Please quantify this risk, e.g., by evaluating models excluded from annotation or reporting agreement between annotators and gold.
  4. [Footnote 2, Figure 1, §5.1] Footnote 2 states the source data is Chinese and 'we use MiGUE-Pipeline to construct the Chinese dataset', yet Figure 1 shows an English example. The paper does not state the language of the evaluated instances. This ambiguity affects the interpretation of all results. Please clarify the language of the benchmark and discuss any language-specific effects.
  5. [§4.2] §4.2 states that event pairs 'that the effective LLMs (such as GPT-4o) initially misclassify' are prioritized to increase difficulty. This selection bias may inflate difficulty and reduce representativeness. Report the fraction of such pairs and consider ablating the selection.
minor comments (3)
  1. [Table 4] Table 4 is very dense; consider splitting or adding vertical separators between granularities to improve readability.
  2. [§5] The paper does not report human performance on any task. Human baselines would help interpret near-chance scores, especially for Causal Graph.
  3. [§3.4] The Leiden clustering hyperparameters and validation of the resulting clusters are not described.

Circularity Check

0 steps flagged

No derivation-level circularity; the benchmark construction is self-contained, with LLM-annotation overlap a validity concern rather than a definitional reduction.

full rationale

MiGUE-Bench does not make a first-principles derivation claim; it presents a benchmark-construction pipeline and empirical evaluation. The gold labels are produced by LLMs in MiGUE-Pipeline (e.g., §3.2 'we make an effective LLM (i.e., DeepSeek-V3.2) generate a candidate set of event triggers'; §3.3 majority voting with GPT-4o, DeepSeek-V3.2, and Qwen3-Max, with Claude-4.5-Opus as arbiter), and the same model families are later evaluated in Table 4. This creates a genuine annotation-bias / self-agreement confound, and §4.5's assertion of 'manual review on all the instances' is not supported by the 200- and 100-sample manual checks reported in §3.2 and §3.3. However, none of these issues makes a stated prediction equal to its input by construction: no parameter is fitted from a subset and then used to predict a closely related quantity, no load-bearing result is imported from the authors' prior work, and the causal-graph and temporal-order gold standards are algorithmically constructed from explicit rules (§3.3 propagate-via-coreference; §4.3.2 maximal paths in causal DAGs). The central low-score findings are not forced by the construction; they could in principle have been different. Under the stated circularity criteria, the derivation chain is self-contained; the annotation-model overlap is a benchmark-validity and reporting concern, not a definitional circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim depends on the quality of automatically generated labels and on several hand-set thresholds in the pipeline; no new physical or formal entities are introduced. The most honest measure of the paper's contribution is therefore the validation strength of its LLM-based annotation assumptions.

free parameters (3)
  • temporal_proximity_window = 1 month
    Cross-document coreference candidates are restricted to documents within a one-month timestamp window (Section 3.3). The window size directly determines which relations can be discovered, and no sensitivity analysis is given.
  • semantic_affinity_threshold = unspecified
    Event mentions must exceed a semantic similarity threshold to be considered candidate coreferents (Section 3.3). The threshold is never quantified, yet it controls the density of the relation network.
  • document_filtering_criteria = LLM judge thresholds unspecified
    Documents are retained only if GPT-4o judges them to satisfy factual informativeness, text quality, and event density criteria (Section 3.1). No scoring rubrics or thresholds are reported, so the retained corpus depends on unstated hand-set criteria.
axioms (4)
  • domain assumption LLM self-correction against retrieved failure examples yields near-human annotation quality.
    Section 3.2 accepts event triggers after retrieval-augmented self-correction with only 200-sample manual validation (precision 0.87). Error types can propagate if the retrieval exemplars are biased.
  • domain assumption Coreference transitivity justifies relation propagation across documents.
    Section 3.3 'Propagate-via-Coreference' applies formal transitivity rules (e.g., coreference plus cause implies cause). This assumes coreference labels are accurate enough that propagation does not amplify mistakes.
  • domain assumption Majority voting by three LLMs yields ground-truth relation labels.
    Section 3.3 accepts relations when GPT-4o, DeepSeek-V3.2, and Qwen3-Max agree, escalating disagreements to Claude-4.5-Opus. External human gold validation is reported on only 100 samples.
  • domain assumption LLM-generated future options and causal-chain distractors are exhaustive and unambiguous.
    Section 4.4 and 4.3.2 generate correct answers and adversarial distractors using LLMs. Options may be eliminable by surface cues, and the true future event is not guaranteed to be the only non-distracting choice.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language Models." pith.science (2026). https://pith.science/paper/B6MOSBLU

@misc{pith2026260727654,
  author       = {Pith},
  title        = {Pith review of: From Single- to Cross-Document: Benchmarking Multi-Granularity Event Analysis of Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6MOSBLU}},
  note         = {Machine review of arXiv:2607.27654}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Event analysis is an essential and fundamental direction of information extraction, involving various event-centric tasks at different granularity of documents. While large language models (LLMs) have preliminarily achieved promising performance in part of these tasks individually, their capability in event analysis still lacks comprehensive understanding due to restricted document granularity, task designs, and data source of existing benchmarks. To address these limitations, we introduce MiGUE-Bench, a systematic benchmark for assessing the performance of LLMs in multi-granularity event analysis. To support large-scale evaluation, we first develop an LLM-driven self-correcting annotation framework called MiGUE-Pipeline, enabling scalable acquisition of high-quality source data of events with automatic labels. Then, we design four core tasks in our benchmark, i.e., event detection, relation reasoning, structure induction, and future prediction, to probe model competence at different levels, from atomic event details to complex cross-document narratives. Extensive experiments on state-of-the-art LLMs and retrieval-augmented generation (RAG) methods delineate the current capability boundary and identify critical deficiencies, providing insights into the future improvement of LLMs in challenging event analysis tasks.

Figures

Figures reproduced from arXiv: 2607.27654 by Guannan Li, Jie Zou, Jinjie Qiu, Ke Qin, Lan Wang, Pei Ke, Shuai Shao, Tao Tian, Tao Wen, Xu Han.

Figure 1
Figure 1. Figure 1: Overview of MiGUE-Pipeline and MiGUE-Bench. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Accuracy on MiGUE-Reasoning (Subevent) with [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 9 linked inside Pith

  1. [1]

    Anthropic. 2025. Introducing Claude Haiku 4.5. https://www.anthropic.com/ news/claude-haiku-4-5. Accessed: 2026-04-25. SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia. Tao Wen et al

    2025

  2. [2]

    Anthropic. 2025. Introducing Claude Opus 4.5. https://www.anthropic.com/ news/claude-opus-4-5. Accessed: 2026-04-25

    2025

  3. [3]

    Michael Bugert, Nils Reimers, and Iryna Gurevych. 2021. Generalizing Cross- Document Event Coreference Resolution Across Multiple Corpora.Computa- tional Linguistics47, 3 (2021), 575–614

    2021

  4. [4]

    Chunkit Chan, Cheng Jiayang, Weiqi Wang, Yuxin Jiang, Tianqing Fang, Xin Liu, and Yangqiu Song. 2024. Exploring the Potential of ChatGPT on Sentence Level Relations: A Focus on Temporal, Causal, and Discourse Relations. InFindings of the Association for Computational Linguistics: EACL 2024. 684–721

    2024

  5. [5]

    Meiqi Chen, Yubo Ma, Kaitao Song, Yixin Cao, Yan Zhang, and Dongsheng Li

  6. [6]

    Muhao Chen, Hongming Zhang, Qiang Ning, Manling Li, Heng Ji, Kathleen McKeown, and Dan Roth. 2021. Event-centric natural language processing. In Proceedings of the 59th Annual Meeting of the Association for Computational Lin- guistics and the 11th International Joint Conference on Natural Language Processing: Tutorial Abstracts. 6–14

    2021

  7. [7]

    Yuxuan Chen, Dewen Guo, Sen Mei, Xinze Li, Hao Chen, Yishan Li, Yixuan Wang, Chaoyue Tang, Ruobing Wang, Dingjun Wu, Yukun Yan, Zhenghao Liu, Shi Yu, Zhiyuan Liu, and Maosong Sun. 2025. UltraRAG: A Modular and Automated Toolkit for Adaptive Retrieval-Augmented Generation.CoRRabs/2504.08761 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

    Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng, and Jun Zhao. 2015. Event extraction via dynamic multi-pooling convolutional neural networks. InProceed- ings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 167–176

    2015

  9. [9]

    Milind Choudhary and Xinya Du. 2024. QAEVENT: Event Extraction as Question- Answer Pairs Generation. InFindings of the Association for Computational Lin- guistics: EACL 2024. 1860–1873

    2024

  10. [10]

    DeepSeek-AI. 2025. DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models.CoRRabs/2512.02556 (2025)

    Pith/arXiv arXiv 2025

  11. [11]

    Doddington, Alexis Mitchell, Mark A

    George R. Doddington, Alexis Mitchell, Mark A. Przybocki, Lance A. Ramshaw, Stephanie M. Strassel, and Ralph M. Weischedel. 2004. The Automatic Content Extraction (ACE) Program - Tasks, Data, and Evaluation. InProceedings of the Fourth International Conference on Language Resources and Evaluation. 837–840

    2004

  12. [12]

    Chuang Fan, Daoxing Liu, Libo Qin, Yue Zhang, and Ruifeng Xu. 2022. Towards Event-level Causal Relation Identification. InSIGIR ’22: The 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1828–1833

    2022

  13. [13]

    John Foley, Michael Bendersky, and Vanja Josifovski. 2015. Learning to Extract Local Events from the Web. InProceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval. 423–432

    2015

  14. [14]

    Jie Gong, Biaoshuai Zheng, and Qiwang Hu. 2025. EventRelBench: A Comprehen- sive Benchmark for Evaluating Event Relation Understanding in Large Language Models. InFindings of the Association for Computational Linguistics: EMNLP 2025. 9084–9099

    2025

  15. [15]

    Google. 2025. Gemini 3: A new era of intelligence. https://blog.google/products- and-platforms/products/gemini/gemini-3/. Accessed: 2026-04-25

    2025

  16. [16]

    Yong Guan, Hao Peng, Xiaozhi Wang, Lei Hou, and Juanzi Li. 2024. OpenEP: Open-Ended Future Event Prediction.CoRRabs/2408.06578 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. 2025. LightRAG: Simple and Fast Retrieval-Augmented Generation. InFindings of the Association for Computational Linguistics: EMNLP 2025. 10746–10761

    2025

  18. [18]

    Yu Hong, Tongtao Zhang, Tim O’Gorman, Sharone Horowit-Hendler, Heng Ji, and Martha Palmer. 2016. Building a Cross-document Event-Event Relation Corpus. InProceedings of the 10th Linguistic Annotation Workshop held in conjunction with ACL 2016. 1–6

    2016

  19. [19]

    Zhilei Hu, Zixuan Li, Xiaolong Jin, Long Bai, Jiafeng Guo, and Xueqi Cheng

  20. [20]

    Kuan-Hao Huang, I-Hung Hsu, Tanmay Parekh, Zhiyu Xie, Zixuan Zhang, Prem Natarajan, Kai-Wei Chang, Nanyun Peng, and Heng Ji. 2024. TextEE: Benchmark, Reevaluation, Reflections, and Future Challenges in Event Extraction. InFindings of the Association for Computational Linguistics: ACL 2024. 12804–12825

    2024

  21. [21]

    Zijie Huang, Jeehyun Hwang, Junkai Zhang, Jinwoo Baik, Weitong Zhang, Do- minik Wodarz, Yizhou Sun, Quanquan Gu, and Wei Wang. 2024. Causal graph ode: Continuous treatment effect modeling in multi-agent dynamical systems. In Proceedings of the ACM Web Conference 2024. 4607–4617

    2024

  22. [22]

    Kimi Team. 2025. Kimi K2: Open Agentic Intelligence.CoRRabs/2507.20534 (2025)

    Pith/arXiv arXiv 2025

  23. [23]

    Yu, and Hao Fei

    Bobo Li, Xudong Han, Jiang Liu, Yuzhe Ding, Liqiang Jing, Zhaoqi Zhang, Jinheng Li, Xinya Du, Fei Li, Meishan Zhang, Min Zhang, Aixin Sun, Philip S. Yu, and Hao Fei. 2025. Event Extraction in Large Language Model.CoRRabs/2512.19537 (2025)

    arXiv 2025

  24. [24]

    Manling Li, Sha Li, Zhenhailong Wang, Lifu Huang, Kyunghyun Cho, Heng Ji, Jiawei Han, and Clare Voss. 2021. The Future is not One-dimensional: Complex Event Schema Induction by Graph Modeling for Event Prediction. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 5203–5215

    2021

  25. [25]

    Rui Li, Wenlin Zhao, Cheng Yang, and Sen Su. 2022. A Dual-Expert Framework for Event Argument Extraction. InThe 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1110–1121

    2022

  26. [26]

    Jinzhi Liao, Xiang Zhao, Xinyi Li, Lingling Zhang, and Jiuyang Tang. 2021. Learn- ing Discriminative Neural Representations for Event Detection. InThe 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 644–653

    2021

  27. [27]

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-Eval: NLG Evaluation using Gpt-4 with Better Human Alignment. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2511–2522

    2023

  28. [28]

    Llama Team. 2024. The Llama 3 Herd of Models.CoRRabs/2407.21783 (2024)

    Pith/arXiv arXiv 2024

  29. [29]

    Chenwei Lou, Jun Gao, Changlong Yu, Wei Wang, Huan Zhao, Weiwei Tu, and Ruifeng Xu. 2022. Translation-Based Implicit Annotation Projection for Zero-Shot Cross-Lingual Event Argument Extraction. InThe 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2076–2081

    2022

  30. [30]

    Yunshan Ma, Chenchen Ye, Zijian Wu, Xiang Wang, Yixin Cao, and Tat-Seng Chua. 2023. Context-aware event forecasting via graph disentanglement. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 1643–1652

    2023

  31. [31]

    Yunshan Ma, Chenchen Ye, Zijian Wu, Xiang Wang, Yixin Cao, Liang Pang, and Tat-Seng Chua. 2023. Structured, Complex and Time-complete Temporal Event Forecasting.CoRRabs/2312.01052 (2023)

    Pith/arXiv arXiv 2023

  32. [32]

    Anne-Lyse Minard, Manuela Speranza, Eneko Agirre, Itziar Aldabe, Marieke Van Erp, Bernardo Magnini, German Rigau, and Ruben Urizar. 2015. Semeval- 2015 task 4: Timeline: Cross-document event ordering. Inproceedings of the 9th International Workshop on Semantic Evaluation (SemEval 2015). 778–786

    2015

  33. [33]

    Paramita Mirza and Sara Tonelli. 2014. An Analysis of Causality between Events and its Relation to Temporal Information. InProceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers. 2097– 2106

    2014

  34. [34]

    Nishanth Nakshatri, Siyi Liu, Sihao Chen, Dan Roth, Dan Goldwasser, and Daniel Hopkins. 2023. Using LLM for improving key event discovery: Temporal-guided news stream clustering with event summaries. InFindings of the Association for Computational Linguistics: EMNLP 2023. 4162–4173

    2023

  35. [35]

    Thien Huu Nguyen, Kyunghyun Cho, and Ralph Grishman. 2016. Joint event extraction via recurrent neural networks. InProceedings of the 2016 conference of the North American chapter of the association for computational linguistics: human language technologies. 300–309

    2016

  36. [36]

    Thien Huu Nguyen and Ralph Grishman. 2018. Graph Convolutional Networks With Argument-Aware Pooling for Event Detection. InProceedings of the Thirty- Second AAAI Conference on Artificial Intelligence. 5900–5907

    2018

  37. [37]

    OpenAI. 2026. OpenAI GPT-5 System Card.CoRRabs/2601.03267 (2026)

    Pith/arXiv arXiv 2026

  38. [38]

    Qwen Team. 2025. Qwen3 Technical Report.CoRRabs/2505.09388 (2025)

    Pith/arXiv arXiv 2025

  39. [39]

    Rashmi Sankepally. 2019. Event Information Retrieval from Text. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval(Paris, France)(SIGIR’19). 1447

    2019

  40. [40]

    Lei Sha, Feng Qian, Baobao Chang, and Zhifang Sui. 2018. Jointly Extracting Event Triggers and Arguments by Dependency-Bridge RNN and Tensor-Based Argument Interaction. InProceedings of the Thirty-Second AAAI Conference on Artificial Intelligence, (AAAI-18), the 30th innovative Applications of Artificial Intelligence (IAAI-18), and the 8th AAAI Symposium ...

    2018

  41. [41]

    Zhengwei Tao, Zhi Jin, Yifan Zhang, Xiancai Chen, Haiyan Zhao, Jia Li, Bin Liang, Chongyang Tao, Qun Liu, and Kam-Fai Wong. 2025. A Comprehensive Evaluation on Event Reasoning of Large Language Models. InThe Thirty-Ninth AAAI Conference on Artificial Intelligence. 25273–25281

    2025

  42. [42]

    Team GLM. 2024. ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools.CoRRabs/2406.12793 (2024)

    Pith/arXiv arXiv 2024

  43. [43]

    Vincent A Traag, Ludo Waltman, and Nees Jan Van Eck. 2019. From Louvain to Leiden: guaranteeing well-connected communities.Scientific reports9, 1 (2019), 1–12

    2019

  44. [44]

    Xiaozhi Wang, Yulin Chen, Ning Ding, Hao Peng, Zimu Wang, Yankai Lin, Xu Han, Lei Hou, Juanzi Li, Zhiyuan Liu, Peng Li, and Jie Zhou. 2022. MAVEN-ERE: A Unified Large-scale Dataset for Event Coreference, Temporal, Causal, and Subevent Relation Extraction. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 926–941

    2022

  45. [45]

    Xiaozhi Wang, Ziqi Wang, Xu Han, Wangyi Jiang, Rong Han, Zhiyuan Liu, Juanzi Li, Peng Li, Yankai Lin, and Jie Zhou. 2020. MAVEN: A Massive General Domain Event Detection Dataset. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 1652–1671. From Single- to Cross-Document: Benchmarking Multi-Granularity Event ...

    2020

  46. [46]

    Zairun Yang, Yilin Wang, Zhengyan Shi, Yuan Yao, Lei Liang, Keyan Ding, Emine Yilmaz, Huajun Chen, and Qiang Zhang. 2025. EventRAG: Enhancing LLM Generation with Event Knowledge Graphs. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 16967–16979

    2025

  47. [47]

    Zhihan Zhang, Yixin Cao, Chenchen Ye, Yunshan Ma, Lizi Liao, and Tat-Seng Chua. 2024. Analyzing Temporal Complex Events with Large Language Models? A Benchmark towards Temporal, Long Context Understanding. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 1588–1606

    2024

  48. [2024]

    In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Improving large language models in event relation logical prediction. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 9451–9478

  49. [2025]

    InProceedings of the 31st International Conference on Computational Linguistics

    Large Language Model-Based Event Relation Extraction with Rationales. InProceedings of the 31st International Conference on Computational Linguistics. 7484–7496

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.