Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Causal Graph based Event Reasoning using Semantic Relation Experts

T0 review · 3 major / 5 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 A collaborative debate among four semantic-relation experts produces more accurate causal event graphs than direct prompting, pairwise classification, or non-collaborative experts, and the resulting graphs improve event forecasting…

desk verdict Solid causal-graph generation result with human-labeled evaluation, but the EEL downstream claim overreaches because it never measures prediction accuracy. read the letter →

arxiv 2506.06910 v1 pith:ZQ6RVTGA submitted 2025-06-07 cs.AI

Mahnaz Koupaee , Xueying Bai , Mudan Chen , Greg Durrett , Nathanael Chambers , Niranjan Balasubramanian This is my paper
classification cs.AI
keywords causaleventgraphsmulti-agentdebatesemanticrelationexpertsreasoningexplainablelikelihoodforecastingnextpredictionlargelanguagemodels
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

This paper tries to establish that an explicit causal event graph—not just pairwise co-occurrence statistics—is a useful backbone for LLM event reasoning, and that the best way to build such a graph is to have four specialised agents argue about it. The agents focus on temporal order, shared-entity discourse links, preconditions, and commonsense, exchange views for several rounds, and a final judge consolidates their edges; on the CRAB benchmark this collaborative process reaches 79.27 graph-level balanced accuracy with GPT-4o, clearly above direct prompting (70.86) and exhaustive pairwise classification (73.93). The same graphs, used without any fine-tuning, drive a new explainable-event-likelihood task, event forecasting, and next-event prediction, where the graph-based method beats few-shot GPT-4 baselines and competes with models trained specifically on those tasks. The stakes are that event reasoning can become simultaneously more accurate and more explainable: every likelihood judgment can point to a causal chain in a global graph.

What carries the argument

The causal event graph is the central object: nodes are events extracted from a news article, directed edges are causal links. The generator that carries the argument is a multi-agent debate protocol: four LLM experts (temporal, discourse, precondition, commonsense) each identify candidate edges while attending to their own semantic relation, exchange and critique one another's lists in rounds, and a causality judge reconciles disagreements and adds transitive closure. The downstream machinery is graph insertion: to judge a query event likely, the model must place it in the graph by naming its causes and effects, and any path through the placed event is read as the explanation. The design rests on the premise that these four relations provide complementary evidence—temporal order is necessary but not sufficient, shared entities and preconditions scope the search, and commonsense supplies unstated links—so that errors from one aspect can be corrected by the others.

What would settle it

Replace the GPT-4 primitive causal question with pooled judgments from many human annotators on all 520 news/query pairs and recompute Causality, Informativeness, and Coherence; if the human-based win rates do not reproduce CGEL's advantage over the one-shot and Reflexion baselines, or differ from GPT-4 by more than the reported margin, the central utility claim fails.

Watch

Extended reading notes

Core claim

The central claim is that a multi-expert debate, rather than a single model pass or pairwise enumeration, is the right way to get LLMs to produce global causal graphs. Four relation experts—temporal, discourse, precondition, and commonsense—each propose causal edges from their own angle, read one another's arguments over up to three rounds, and a causality judge compiles the final graph, including transitive links. The paper reports that this procedure outperforms every baseline on the CRAB causal-graph benchmark, and that ablations removing either the debate or any one expert degrade accuracy. Downstream, the method defines CGEL: a query event is likely if it can be inserted into the graph as a cause or effect of existing events, and the causal path through the query serves as the explanation. On the new 520-pair Explainable Event Likelihood test set, CGEL beats one-shot and Reflexion baselines on causality, informativeness, and coherence; on ForecastQA yes/no forecasting it reaches 62.7% accuracy versus 51.3% for direct GPT-4 answering, and on the next-event cloze task it reaches 61% with context versus 38% for the one-shot baseline.

Load-bearing premise

The central utility claim assumes that GPT-4's yes/no answers to "Can event X cause event Y?" faithfully measure whether an explanation is good, and that explanation quality can stand in for correctness of the likelihood prediction; this was checked against humans only on a small subset, and no gold likelihood labels exist.

Editorial extensions

If this is right

  • If the graph-generation results hold, LLM event reasoning can be made inspectable: every likelihood prediction carries a causal chain that can be checked edge by edge.
  • The method offers a no-fine-tuning recipe that is competitive on forecasting and next-event prediction, suggesting explicit causal structure can substitute for some task-specific training data.
  • The ablations imply that separation of concerns and communication, not model scale alone, drive the improvement: deleting any of the four experts or silencing the debate lowers accuracy.
  • The new Explainable Event Likelihood task and its three metrics give a reusable template for evaluating whether an explanation actually supports a likelihood judgment.

Reading between the lines

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

  • A natural extension not tested in the paper is adding a fifth expert for a different causal dimension (for example, statistical or motivational causes); the paper's claim that four experts capture all needed subtleties would be falsified if a fifth expert consistently shifts edges toward human judgment.
  • Because the EEL metrics treat GPT-4's answer to "can event X cause event Y?" as the gold standard, the reported wins are, strictly, alignment with GPT-4's causality notion; full-set human annotation could change the ranking even though a 10-article subset showed agreement.
  • The binary insertion rule discards graded causal strength; incorporating the graded causality scores noted in the paper's limitations could yield calibrated likelihoods and more nuanced explanations.
  • Since 25% of debates terminate early, a dynamic stopping rule based on inter-expert agreement could cut API cost while preserving most of the accuracy gain; the paper measures costs only for fixed-round debates.
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.

Referee Report

3 major / 5 minor

Summary. The paper proposes a collaborative multi-agent method for causal event graph generation in which four LLM-based experts (temporal, discourse, precondition, and commonsense) debate over multiple rounds and a judge consolidates the final graph. The method is evaluated intrinsically on the CRAB dataset, where it outperforms direct, pairwise, and non-collaborative baselines for both GPT-4o and Llama-70B. The paper then applies these causal graphs to a new Explainable Event Likelihood (EEL) task, to ForecastQA, and to a narrative-cloze next-event prediction task, reporting better explanation quality than one-shot and Reflexion baselines and competitive accuracy with specialized models without fine-tuning.

Significance. If the central claims held, the work would be a useful step toward making LLM event reasoning more causally grounded and interpretable. The intrinsic CRAB experiments are carefully designed: the collaborative approach shows consistent gains over meaningful baselines, the ablations demonstrate the contribution of each expert, and the cost analysis is transparent. The main significance is currently limited by the EEL evaluation, which does not measure whether the likelihood prediction itself is correct, and by the use of GPT-4 as judge for explanations produced by GPT-4-based systems. The downstream ForecastQA and narrative-cloze results are real accuracy measurements, but they are reported on small subsets without statistical support.

major comments (3)
  1. [Sec. 4.1-4.2, Appendix B.3.1] The EEL experiment never scores the CGEL likelihood decision itself. The system predicts that a query event is likely iff it can be inserted into the generated causal graph (Sec. 4.1), but no gold likelihood labels are used in the evaluation; Appendix B.3.1 states that such labels were abandoned after low inter-annotator agreement. The three metrics (Causality, Informativeness, Coherence) are computed from GPT-4's answers to the primitive question 'Can EVENT1 cause EVENT2?' over adjacent pairs in the explanation chain (Appendix B.3), so they measure internal causal coherence of the explanation, not whether the queried event is actually likely. A system can insert an unlikely event into the graph and still produce a chain that scores high on all three metrics. Therefore the Introduction's claim of 'significantly better predictions' (Sec. 1) and the abstract's utility claim are not supported by the EEL results. The authors should either add a human-annotated or otherwise gold-labeled likelihood evaluation on at least a subset, or explicitly restrict the EEL claim to explanation quality rather than prediction accuracy.
  2. [Sec. 4.2.1, Table 4, Appendix B.5.1] The EEL headline comparisons use GPT-4 as both the generator and the evaluator of causal relations, which creates a same-family evaluator risk that is acknowledged in the Limitations section. The human validation in Appendix B.5.1 covers only 10 articles and reports no inter-annotator agreement statistic, so it is hard to gauge how reliable the GPT-4 judgments are. To support the pairwise win rates in Table 4, the paper should report agreement statistics (e.g., Cohen's kappa) on a larger human sample, and ideally provide human evaluation of all three metrics on a random subset rather than a single aggregate comparison.
  3. [Sec. 4.3, Table 5 and Sec. 4.4, Table 6] The downstream claims of competitiveness rest on small subsets: 25% of ForecastQA and 10% of the narrative-cloze test set. No confidence intervals or significance tests are reported, and it is not stated how the subsets were sampled or whether they are representative of the full test sets. Given the small numbers, the differences in Table 5 (e.g., 62.7 for CGEL versus 63.1 for BERT base and 67.4 for BERT large) may not be robust. The paper should report the exact subset composition, bootstrap confidence intervals, and significance tests, or evaluate on the full test sets if feasible.
minor comments (5)
  1. [Sec. 2.1] The heading 'Conditional Expert' is inconsistent with the term 'precondition expert' used throughout the rest of the paper and in the prompt tables; please unify the terminology.
  2. [Appendix B] The item numbering is duplicated: '3) Event placement' and '3) Explanation generation' both appear. Please renumber the steps sequentially.
  3. [Table 22, Appendix B.5.1] Please report the number of event pairs used in the human evaluation and the inter-annotator agreement, not just the aggregate percentages, so readers can assess the reliability of the comparison.
  4. [Sec. 4.3] The forecasting formulation uses leaf-node placement only, while the EEL task allows arbitrary insertion into the graph. The reason for this difference is not explained; please clarify why leaf placement is the appropriate decision rule for forecasting.
  5. [Abstract and Introduction] The phrase 'not finetuned' should be hyphenated as 'not fine-tuned' for consistency with standard terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the causal graph generation is scored against human-labeled CRAB data, and the downstream utility claims are checked against gold-labeled ForecastQA and narrative cloze accuracy.

full rationale

The paper's central derivation chain is not circular. The intrinsic evaluation of collaborative causal graph generation uses the CRAB benchmark with human causality labels (Section 3.1, Table 1), providing an external ground truth independent of the proposed method. The downstream utility claims rest on two gold-label evaluations: event forecasting accuracy on ForecastQA (Table 5) and next-event-prediction accuracy on the narrative cloze task (Table 6), both of which measure whether the system's graph-insertion decision matches a known correct answer. The EEL experiments do rely on GPT-4's answers to the primitive causal question for explanation-quality metrics, and the paper's own Limitations section acknowledges that automatic evaluation may exhibit bias toward GPT-4; however, Appendix B.5.1 reports a human evaluation on a subset of the data with similar trends (Table 22), which is an independent check rather than a construction-level equivalence. The use of the Devare et al. schemas to curate EEL test instances is data curation, not a load-bearing self-citation of the paper's own result, and no fitted parameter or asserted uniqueness theorem is invoked to force the reported outcomes. The absence of gold likelihood labels in the EEL evaluation is a validity concern about what the explanation metrics certify, but it does not make any prediction equivalent to its inputs by definition. Therefore, no circular step is evidenced.

Assumptions & free parameters 2 free parameters · 6 assumptions · 0 invented entities

The central claim rests on domain assumptions about the semantics of causality (temporal precedence, entity sharing, counterfactual dependence, commonsense mediation) and on the reliability of LLM-generated events and LLM-based evaluation. No new physical or mathematical entities are introduced. The free parameters are the CRAB labeling threshold and the maximum debate rounds, both hand-set choices.

free parameters (2)
  • CRAB causal threshold = 50
    Pairs with causality score above 50 are labeled causal, at or below 50 non-causal. This binary threshold is chosen from the dataset's 0-100 scale and determines the gold labels for intrinsic evaluation (Section 3.1).
  • Maximum debate rounds = 3
    Experts discuss for up to 3 rounds before the judge consolidates. The round count was chosen based on the plateau in Figure 3; it is a hand-set hyperparameter, not fitted by optimization.
assumptions (6)
  • domain assumption Temporal precedence is a necessary condition for causality: if ei causes ej then ei precedes ej.
    Used by the temporal expert to filter candidate causal pairs (Section 2.1).
  • domain assumption Events sharing an entity are more likely to be causally related, providing a scoping heuristic.
    Discourse expert uses shared participants to identify candidate causal links (Section 2.1).
  • domain assumption If removing an event ei would make ej irrelevant, then ei is a precondition and likely causal for ej.
    Precondition expert evaluates dependence by counterfactual removal (Section 2.1).
  • domain assumption LLM-generated event extractions from the news text are accurate enough for graph construction.
    The system depends on GPT-4's extracted events E as graph nodes, with no verification against gold event annotations (Appendix B, step 1).
  • domain assumption CRAB human causality scores are a reliable ground truth for causal graph evaluation.
    CRAB scores are used to define gold causal/non-causal pairs (Section 3.1).
  • domain assumption GPT-4's answers to the primitive question 'Can event EVENT1 cause event EVENT2?' are a valid proxy for human causal judgment in explanation evaluation.
    The EEL evaluation metrics are computed from GPT-4's pairwise causal judgments; human validation was performed on a subset only (Appendix B.5.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal Graph based Event Reasoning using Semantic Relation Experts." pith.science (2026). https://pith.science/paper/ZQ6RVTGA

@misc{pith2026250606910,
  author       = {Pith},
  title        = {Pith review of: Causal Graph based Event Reasoning using Semantic Relation Experts},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZQ6RVTGA}},
  note         = {Machine review of arXiv:2506.06910}
}
read the original abstract

Understanding how events in a scenario causally connect with each other is important for effectively modeling and reasoning about events. But event reasoning remains a difficult challenge, and despite recent advances, Large Language Models (LLMs) still struggle to accurately identify causal connections between events. This struggle leads to poor performance on deeper reasoning tasks like event forecasting and timeline understanding. To address this challenge, we investigate the generation of causal event graphs (e.g., A enables B) as a parallel mechanism to help LLMs explicitly represent causality during inference. This paper evaluates both how to generate correct graphs as well as how graphs can assist reasoning. We propose a collaborative approach to causal graph generation where we use LLMs to simulate experts that focus on specific semantic relations. The experts engage in multiple rounds of discussions which are then consolidated by a final expert. Then, to demonstrate the utility of causal graphs, we use them on multiple downstream applications, and also introduce a new explainable event prediction task that requires a causal chain of events in the explanation. These explanations are more informative and coherent than baseline generations. Finally, our overall approach not finetuned on any downstream task, achieves competitive results with state-of-the-art models on both forecasting and next event prediction tasks.

Figures

Figures reproduced from arXiv: 2506.06910 by the authors.

Figure 1
Figure 1. A causal graph consisting of events as nodes and causal relations as directed edges. 2018a,b; Koupaee et al., 2021), and controllable generation with larger language models (Gunjal and Durrett, 2023; Tang et al., 2023; Dror et al., 2023). While recent work has advanced event rea￾soning capabilities overall, particularly in event prediction, most work has relied primarily on dis￾tributional relations between events c… view at source ↗
Figure 2
Figure 2. Collaborative causal graph generation with relation experts. Initially, the experts start communicating with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The effect of collaboration rounds on perfor [PITH_FULL_IMAGE:figures/full_fig_p025_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Overview of causal graph-based event likelihood prediction approach. [PITH_FULL_IMAGE:figures/full_fig_p026_4.png]
Figure 5
Figure 5. Figure 5: All the prompts used for models evaluated for event reasoning tasks. The [PITH_FULL_IMAGE:figures/full_fig_p028_5.png]
Figure 6
Figure 6. Figure 6: Performance of systems based on different selection heuristics. Each bar shows the difference in [PITH_FULL_IMAGE:figures/full_fig_p030_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages

  1. [1]

    Temporal expert (you), responsible for taking into account the temporal relations as prerequisites for a causal relation

    work page

  2. [3]

    Forecastqa: A question answering challenge for event forecasting with temporal text data. In Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 4636– 4650. Xiaomeng Jin, Manling Li, and Heng Ji. 2022. Event schem...

    work page 2022

  3. [4]

    You can see your previous responses as well as other experts responses

    Discourse expert, responsible for identifying whether events are sharing entities that can lead to identifying causal links. You can see your previous responses as well as other experts responses. You can see the list of all identified pairs by all experts. Continue the discussion with other evaluator experts, talk to them and state why you agree/disagree...

    work page

  4. [5]

    Zhiyi Luo, Yuchen Sha, Kenny Q Zhu, Seung-won Hwang, and Zhongyuan Wang

    Can large language models build causal graphs? arXiv preprint arXiv:2303.05279. Zhiyi Luo, Yuchen Sha, Kenny Q Zhu, Seung-won Hwang, and Zhongyuan Wang. 2016. Commonsense causal reasoning between short texts. In Fifteenth international conference on the principles of knowl- edge representation and reasoning. Zhao Mandi, Shreeya Jain, and Shuran Song. 2024...

    arXiv 2016

  5. [6]

    The gap is very small for the causality whereas the longer chains lead to larger gap between the CGEL and the baseline in terms of informativeness and coherence

    Each bar shows the difference in performance of selection criteria on the specified metric. The gap is very small for the causality whereas the longer chains lead to larger gap between the CGEL and the baseline in terms of informativeness and coherence. However, regardless of the selection heuristic, CGEL outperforms the baseline in all 3 dimensions. B.5....

    work page 2021

  6. [7]

    arXiv preprint arXiv:2311.04284

    Crab: Assessing the strength of causal rela- tionships between real-world events. arXiv preprint arXiv:2311.04284. Evan Sandhaus. 2008. The new york times annotated corpus. Linguistic Data Consortium, Philadelphia, 6(12):e26752. Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019. Distilbert, a distilled version of bert: smaller, faster, c...

    arXiv 2008

  7. [12]

    Discourse expert (you), responsible for identifying whether events are sharing entities that can lead to identifying causal links

    work page

  8. [15]

    You can see your previous responses as well as other experts responses

    Precondition expert, responsible for checking how one event would affect the other event, mainly if removing one of the events would also lead to making the other one irrelevant. You can see your previous responses as well as other experts responses. You can see the list of all identified pairs by all experts. Continue the discussion with other evaluator ...

    work page

Show all 39 references
  1. [16]

    Precondition expert (you), responsible for checking how one event would affect the other event, mainly if removing one of the events would also lead to making the other one irrelevant

  2. [19]

    You can see your previous responses as well as other experts responses

    Discourse expert, responsible for identifying whether events are sharing entities that can lead to identifying causal links. You can see your previous responses as well as other experts responses. You can see the list of all identified pairs by all experts. Continue the discus...

  3. [20]

    Commonsense expert (you), responsible for identifying the missing commonsense bits that can help with identifying whether there is a causal link or not

  4. [23]

    You can see your previous responses as well as other experts responses

    Discourse expert, responsible for identifying whether events are sharing entities that can lead to identifying causal links. You can see your previous responses as well as other experts responses. You can see the list of all identified pairs by all experts. Continue the discus...

  5. [24]

    Temporal expert, responsible for taking into account the temporal relations as prerequisites for a causal relation

  6. [25]

    Precondition expert, responsible for checking how one event would affect the other event, mainly if removing one of the events would also lead to making the other one irrelevant

  7. [26]

    Commonsense expert, responsible for identifying the missing commonsense bits that can help with identifying whether there is a causal link or not

  8. [27]

    You can see the discussion history and the list of all identified pairs by all experts

    Discourse expert, responsible for identifying whether events are sharing entities that can lead to identifying causal links. You can see the discussion history and the list of all identified pairs by all experts. Discussion history: [MESSAGES] Causal pairs: [CAUSAL PAIRS] Fina...

  9. [28]

    Each causal pair must have a temporal relation such that the cause precedes the effect

    looking into the temporality of the given events and evaluate the causality from the temporal perspective. Each causal pair must have a temporal relation such that the cause precedes the effect

  10. [29]

    Events with causal relation can potentially be sharing some entities

    looking into common entities that are shared among the given events and evaluate the causality from that perspective. Events with causal relation can potentially be sharing some entities

  11. [30]

    Events are dependent if removing one of them (assuming it did not happen) leads to making the other event irrelevant

    looking into the dependence of the given events and evaluate the causality from the dependence perspective. Events are dependent if removing one of them (assuming it did not happen) leads to making the other event irrelevant

  12. [31]

    The missing commonsense can be some intermediate events that can help with identifying the link between the given pair of events

    looking into identifying the commonsense that is relevant to the given events and use them to identify whether there is a causal link or not. The missing commonsense can be some intermediate events that can help with identifying the link between the given pair of events. Place...

  13. [32]

    Event Extraction Given the input text D and the query event eq, the first step is to extract all the events E =< e1, e2, ...en > mentioned in the text. In an in-context learning setup, given D and an extraction prompt pe, we use model M to generate a list of all events that ar...

  14. [33]

    We use our proposed collaborative causal graph generation approach to generate the causal graph of events

    Causal Graph Generation The next step cre- ates a causal graph G(E, R) from the extracted events E as the nodes and edges are causal rela- tions R, where (ei, ej) ∈ R implies that ei is the cause of the effect ej using our proposed collabora- tive approach. We use our proposed...

  15. [34]

    full confidence

    Event placement To assess whether a query event eq is likely in the context of the given events, we need to identify its potential causes and effects in the causal graph. This translates to finding an insertion point within the causal graph. If the query Graph-level Pair-level...

    2022

  16. [38]

    Depositors, mostly technology workers and venture capital-backed com- panies, hurried to withdraw their money this week,

    The technology sector face crisis. Conversation among agents (Round 1): Temporal agent: Based on the temporal aspect of the events presented, I identify potential causal relationships. Specifically, the event in which deposi- tors hurried to withdraw their money (Event 2) is d...

  17. [42]

    The technology sector faces crisis

    The technology sector face crisis. Conversation among agents (Round 2): Temporal agent: In discussing the identified causal pairs, I maintain my stance on the relationship between the events involving the depositors and the collapse of Silicon Valley Bank (SVB). The sequence i...

  18. [43]

    Thus, in summary, I suggest adding: - The causal pair 1,2 as it is defined

    directly contributed to the collapse of SVB (Event 1). Thus, in summary, I suggest adding: - The causal pair 1,2 as it is defined. - The discourse of potential influences of the collapse on the tech sector does not provide a straightforward causal chain. With that being said, ...

  19. [44]

    Silicon Valley Bank (SVB) collapsed

  20. [45]

    Depositors, mostly technology workers and venture capital-backed companies, hurried to withdraw their money this week

  21. [46]

    SVB had specialised in financing start-ups

  22. [47]

    The technology sector face crisis. Conversation among agents (Round 3): Temporal agent: In reviewing the evaluations provided by the different agents, I find a strong consensus regarding the causal relationship between event 2 (Depositors hurried to withdraw their money) and e...

  23. [48]

    EVENT1” cause event “EVENT2

    Explanation generation If eq is placed in graph G(E′, R′), it means that the event is likely and then any path P in the graph that includes eq can be seen as a causal explanation for its likeli- hood, as the path encompasses all previous causes and subsequent effects of eq. B....

    2023

  24. [49]

    Causality The causality metric measures how good the generated explanations are in terms of capturing the cause and effect relations between events that are generated as part of the explanation chain, and might be relevant to the query event eq. Using the responses from the pr...

  25. [50]

    Then, we move in both directions and count the number of correct causal relations until an incorrect rela- tion is met

    Informativeness To measure informativeness, we first locate eq in the explanation chain. Then, we move in both directions and count the number of correct causal relations until an incorrect rela- tion is met. The resulting chain is the longest chain with consecutive positive c...

  26. [51]

    Similar to informativeness, we find the longest sub-chain in which all event pairs are cor- rectly causally connected (based on responses to the primitive causal question)

    Coherence A coherent system is capable of correctly connecting pieces of information based on causal relations to form a clear likelihood ex- planation. Similar to informativeness, we find the longest sub-chain in which all event pairs are cor- rectly causally connected (based...

  27. [52]

    EVENT1” cause event “EVENT2

    This shows that even though the system correctly captured more than half of the casual relations, it completely fails in informativeness with respect to the event we care about (eq). B.3.1 A note on the choice of evaluation metrics It is worth mentioning to note that there doe...

    2023

  28. [424]

    Andrew Gordon, Cosmin Bejan, and Kenji Sagae

    Springer. Andrew Gordon, Cosmin Bejan, and Kenji Sagae. 2011. Commonsense causal reasoning using millions of personal stories. In Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 25, pages 1180–1185. Anisha Gunjal and Greg Durrett. 2023. Drafting event s...

    2011 arXiv

  29. [2003]

    In Corpus linguistics, volume 2003, page 40

    The timebank corpus. In Corpus linguistics, volume 2003, page 40. Lancaster, UK. Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, et al. 2024. Chatdev: Communicative agents for software development. In Proceeding...

    2003

  30. [2021]

    pages 2354–2363, On- line

    ExCAR: Event graph knowledge enhanced explainable causal reasoning. pages 2354–2363, On- line. Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenen- baum, and Igor Mordatch. 2023. Improving factual- ity and reasoning in language models through multia- gent debate. arXiv prepr...

    2023 arXiv

  31. [2023]

    In Proceedings of the First Workshop on Matching From Unstructured and Structured Data (MATCHING 2023), pages 23–32

    Toward consistent and informative event-event temporal relation extraction. In Proceedings of the First Workshop on Matching From Unstructured and Structured Data (MATCHING 2023), pages 23–32. Mahnaz Koupaee, Greg Durrett, Nathanael Chambers, and Niranjan Balasubramanian. 2021...

    2023 arXiv

Pith tools

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