Pith. sign in

REVIEW 2 major objections 6 minor 42 references

ACCESS : A Benchmark for Abstract Causal Event Discovery and Reasoning

T0 review · 2 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper claims that everyday causal knowledge can be organized at an abstraction level and that this organization is both learnable and useful; it introduces ACCESS, a benchmark of 725 abstract events and 1,494 cause–effect pairs built…

desk verdict A genuinely useful new benchmark for abstraction-level causal events, built carefully from GLUCOSE, but the headline QA experiment leaks the gold causal relation and needs a held-out split before the 'enhances QA' claim can be believed. read the letter →

arxiv 2502.08148 v1 pith:ZNBSSXH4 submitted 2025-02-12 cs.AI

Vy Vo , Lizhen Qu , Tao Feng , Yuncheng Hua , Xiaoxi Kang , Songhai Fan , Tim Dwyer , Lay-Ki Soon
show 1 more author
Gholamreza Haffari
This is my paper · ORCID
classification cs.AI
keywords abstractcausaleventseventabstractiondiscoverybenchmarkGLUCOSEreasoningquestionansweringlargelanguagemodels
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

The paper claims that everyday causal knowledge can be organized at an abstraction level and that this organization is both learnable and useful. It introduces ACCESS, a benchmark of 725 abstract events and 1,494 cause–effect pairs built from 9,513 GLUCOSE stories, and reports that adding abstract causal relations as prompt context improves LLM multiple-choice causal QA accuracy by up to 20 percentage points, while LLMs still struggle to identify abstractions or pairwise causal relations on their own. If the central claim holds, abstraction-level causal resources become a practical lever for improving reasoning in question answering, and ACCESS becomes a reusable testbed for measuring progress in abstract causal discovery.

What carries the argument

The central object is the causal graph of event abstractions, where an event abstraction is defined as a tuple ⟨action/state, participant, time, location⟩ shared by all mentions that describe the same event, with time and location optional. Causation is defined counterfactually: x causes y if y temporally follows x directly and y would not commonly occur if x had not occurred. The construction pipeline uses correlation clustering (a PIVOT-style algorithm) with a similarity score combining sentence-embedding cosine similarity and paraphrase likelihood, then human annotation; causal edges come from GLUCOSE's annotated relations and from the PC structure-learning algorithm run on a stories-by-abstractions co-occurrence matrix, all filtered by human annotators.

What would settle it

Check whether any GLUCOSE-QA test story or its paraphrase appears among the 9,513 GLUCOSE stories used to build ACCESS; then rerun the +CG condition with the causal relation replaced by a random or reversed causal pair. If accuracy does not drop, the improvement comes from the hint itself, not from causal knowledge.

Watch

Extended reading notes

Core claim

ACCESS is presented as the first benchmark that pairs corpus-grounded event abstractions with a commonsense causal graph: each node is an abstraction (a cluster of semantically equivalent event mentions, e.g., 'a person needs money'), each edge is a direct cause–effect relation between two abstractions (e.g., needing money leads to getting a job), and every node is grounded in the original story corpus so that statistical causal discovery can run on co-occurrence data. The paper's central claim is that this abstraction-level representation captures causal knowledge that survives paraphrasing and lexical variation, and that it transfers to reasoning: injecting the relevant abstract cause–effect pair into the prompt improves multiple-choice QA for all tested LLMs, with the largest gains on abstract questions where the target event has been generalized. The paper also documents that the same benchmark exposes current limits: automatic clustering of LLM-generated generalizations cannot reproduce the human abstraction level, LLMs achieve only modest pairwise causal discovery F1, and standard statistical structure-learning algorithms recover subgraphs poorly.

Load-bearing premise

The QA gains depend on the untested assumption that the evaluation stories are disjoint from the stories used to build ACCESS, and that giving the model the correct abstract cause–effect relation as a prompt hint is a fair reasoning test rather than answer leakage.

Editorial extensions

If this is right

  • ACCESS can serve as a reusable benchmark for two sub-tasks: abstract event identification and pairwise or full-graph causal discovery at the abstraction level.
  • The +CG result suggests that augmenting prompts with abstract causal knowledge is a straightforward way to improve LLM performance on causal QA without fine-tuning.
  • The benchmark's confounding, mediation, and collider structures allow evaluation of causal inference beyond simple pair classification.
  • LLMs' over-generalization in abstraction identification indicates that controlling abstraction granularity is an open problem, and ACCESS provides ground-truth clusters for testing that skill.
  • The poor performance of statistical structure learning on the co-occurrence representation points to the need for better representations of abstract events.

Reading between the lines

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

  • If the +CG improvement is robust to held-out stories and to control hints, abstraction-level causal graphs could be plugged into retrieval-augmented reasoning pipelines for narrative domains beyond children's stories.
  • The paper leaves implicit that the benchmark's causal graph could be used for Pearl's three rungs of causation (association, intervention, counterfactual); the collider and confounder structures suggest interventional queries could be tested, though the current graph may be too sparse for that.
  • The abstraction definition relies on a fixed ontology and on human judgment of 'same event,' so the benchmark inherits the coverage and biases of GLUCOSE's children's story domain; extending the pipeline to news or scientific text is a direct test of transferability.
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

2 major / 6 minor

Summary. The paper introduces ACCESS, a benchmark for abstract causal event discovery and reasoning, constructed from the GLUCOSE corpus. It proposes a two-phase pipeline: (1) extracting event abstractions by clustering GLUCOSE generalizations with automatic methods followed by human annotation, and (2) discovering causal relations among the resulting 725 abstractions using GLUCOSE annotations, PC-based candidates, and human validation, yielding 1,494 causal pairs. The authors evaluate (a) automatic abstraction with GPT-4o-mini and clustering, (b) pairwise causal discovery by LLMs, (c) statistical structure learning, and (d) the use of ACCESS causal graphs to improve multi-choice QA reasoning on a GLUCOSE-based QA set. The paper reports that automatic abstraction and causal discovery remain challenging, but that adding ACCESS causal information improves QA accuracy and F1 across models.

Significance. If the claims hold, ACCESS is a reusable resource: it is one of the few benchmarks providing corpus-grounded event abstractions together with an abstraction-level causal graph, and the pipeline is documented in enough detail to be reproduced. The paper also honestly documents negative results for LLM-based abstraction and causal discovery, which is useful for the community. However, the central practical claim—that ACCESS's abstract causal knowledge can be leveraged for QA reasoning—is not yet supported because the QA experiments supply the gold causal relation directly in the prompt and do not establish disjointness between the QA stories and the construction corpus. The benchmark construction itself appears careful, with human annotation and explicit quality criteria, and the released code and data strengthen reproducibility.

major comments (2)
  1. [Section 4.3, Appendix E Table 10] The +CG condition inserts the line "A possible cause/effect of the event [abstraction] is [abstraction]" into the prompt, which is exactly the gold abstraction-level causal relation for the queried event. This makes the experiment an oracle evaluation: the model is handed the answer-relevant knowledge, and the paper itself acknowledges in Section 4.3 that for an arbitrary QA dataset this relation would have to be retrieved by first performing abstraction and then mapping to ACCESS. Since Section 4.1 shows that automatic abstraction with GPT-4o-mini and clustering is far from reliable, the QA gains in Table 5 do not support the claim that ACCESS can be leveraged in a realistic pipeline. The authors should evaluate with predicted abstractions/relations, or at minimum include a control condition with randomly selected or unrelated causal relations to show the improvement is specifically attributable to correct abstract causal knowledge.
  2. [Section 4.3, Table 3] The GLUCOSE-QA dataset is described as "constructed from GLUCOSE" (Section 4.3), but the paper never reports a split between the stories used to build ACCESS (9,513 stories, Table 3) and the stories used for GLUCOSE-QA. If the QA stories overlap with the construction corpus, the +CG condition leaks the annotations that produced the answer labels: the supplied abstraction relation is derived from the same corpus-level causal annotations that define the correct answers. The authors should report the degree of overlap and re-run the experiment on a strictly disjoint set, since without this the reported improvements may be an artifact of data leakage rather than evidence of reasoning improvement.
minor comments (6)
  1. [Abstract and Table 3] The abstract states "1, 4K causal pairs" while Table 3 reports 1,494 edges; please use a consistent notation throughout.
  2. [Section 4.1] The text twice uses "ground-true" where "ground-truth" is intended.
  3. [Appendix E, Tables 16 and 17] The example phrase "another thing (that is an ingredient in another thing'" is missing a closing parenthesis; fix the typo.
  4. [Appendix B.2] The term "inter-rate agreement" should be "inter-rater agreement".
  5. [Figure 2] The x-axis labels "5 (4) 7 (6) 16 (15) 19 (18) 45 (48)" are ambiguous; clarify that the first number is nodes and the parenthetical is edges, or use explicit axis labels.
  6. [Table 3] The reported count of 3,956 colliders is surprising for a graph with 725 nodes; please clarify whether this counts collider triples, how it is computed, and why the number exceeds the number of edges.

Circularity Check

1 steps flagged · score 6.0 of 10

The +CG QA experiment supplies the gold causal relation directly in the prompt, so the reported QA gains are an oracle-hint effect rather than evidence for independently usable abstract causal knowledge; the benchmark construction itself is externally grounded.

  1. self definitional [Section 4.3 (Reasoning with Causal Graphs), Appendix E, Tables 10-11]
    "For every story, we create two alternative multi-choice questions about the cause and effect. We extract the sentences in the story as candidate answers. The event appears in the annotated causal pair is thus considered a correct answer. ... (*) This information can help answer the question: A possible <cause/effect> of the event <Input effect/cause event abstraction> is <Input cause/effect event abstraction>."

    The correct answers in GLUCOSE-QA are defined by the annotated causal pair, and the +CG condition inserts that same causal pair as a hint for the queried event. The model is therefore handed the gold relation that determines the answer labels, so the reported gains (e.g., Specific QA accuracy 0.790->0.894 for GPT-4o-mini) measure an oracle-hint effect rather than the benefit of independently accessible abstract causal knowledge. The paper itself notes that for an arbitrary QA dataset the abstraction and relation must be retrieved in two further steps, so the experiment as run does not support the claimed practical use of ACCESS.

full rationale

ACCESS's construction is not circular: the 1,494 causal pairs are derived from GLUCOSE's generalized event mentions plus a documented human-annotation pipeline, and the abstraction/clustering and statistical causal discovery experiments are evaluated against this externally grounded resource. Those experiments mostly report negative results, which is consistent with independent content. The circularity is confined to the QA demonstration in Section 4.3. The +CG prompt line inserts the exact abstraction-level causal relation that defines which candidate sentence(s) are correct, so the improvement is at least partly guaranteed by construction: the model is given the answer-determining relation as a hint. The paper also acknowledges that in a real deployment the relation would need to be retrieved automatically, making the experiment an oracle setting. A further unresolved risk is that GLUCOSE-QA is 'constructed from GLUCOSE' with no described split from the 9,513 GLUCOSE stories used to build ACCESS, so the QA stories and the causal relations used as hints may come from the same documents. Resolving this overlap would be needed before the QA gains can be attributed to reusable abstract knowledge. These issues affect the main positive claim of practical benefit, but not the benchmark resource itself, yielding a partial circularity score of 6.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The benchmark construction depends on several hand-chosen thresholds and subjective annotation decisions; the QA experiment additionally depends on an unstated data split. No new theoretical entities are introduced, only a new dataset and pipeline.

free parameters (5)
  • PIVOT similarity threshold = 0.70
    Hand-chosen threshold for clustering event mentions in Section 3.1 and Appendix C; controls cluster granularity and the final abstraction set.
  • similarity weighting alpha = 0.5
    Equal weight between cosine similarity and paraphrase probability in Eq. (1); arbitrary and directly shapes the clusters.
  • minimum cluster size and similarity post-filter = 10 instances / 0.50 max pairwise similarity
    Appendix C removes clusters with fewer than 10 samples or max pairwise similarity below 50%, affecting the final 725 abstractions.
  • PC test p-value = 0.01
    Used for Chi-squared and G-squared conditional independence tests in Section 3.2 to generate candidate causal pairs.
  • document frequency thresholds for subgraphs = 25, 30, 35, 40, 45
    Appendix D selects subgraphs for statistical structure learning; the choice affects the reported SHD/F1 results.
assumptions (4)
  • domain assumption GLUCOSE's generalized inference rules provide valid event mentions at a consistent level of abstraction.
    The entire pipeline in Section 3.1 uses GLUCOSE's generalized expressions as event mentions and builds clusters from them.
  • domain assumption Counterfactual theory of causation (Lewis) with direct temporal succession and counterfactual dependence is an appropriate definition for event causality.
    Section 2 defines causation according to Lewis; all annotation and evaluation criteria follow this definition.
  • domain assumption Human annotators' commonsense judgments of semantic similarity and causality are treated as ground truth.
    Section 3.1-3.2 relies on 10 and 3 annotators respectively; Krippendorff's alpha of 77.2% is reported, but no adjudication against an external gold standard.
  • domain assumption The PC algorithm's assumptions (causal sufficiency, faithfulness, acyclic underlying graph) hold for the abstraction-level co-occurrence data.
    Section 3.2 and Appendix D apply PC to the co-occurrence matrix; if these assumptions fail, PC candidates are unreliable, though human filtering mitigates this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ACCESS : A Benchmark for Abstract Causal Event Discovery and Reasoning." pith.science (2026). https://pith.science/paper/ZNBSSXH4

@misc{pith2026250208148,
  author       = {Pith},
  title        = {Pith review of: ACCESS : A Benchmark for Abstract Causal Event Discovery and Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZNBSSXH4}},
  note         = {Machine review of arXiv:2502.08148}
}
abstract

Identifying cause-and-effect relationships is critical to understanding real-world dynamics and ultimately causal reasoning. Existing methods for identifying event causality in NLP, including those based on Large Language Models (LLMs), exhibit difficulties in out-of-distribution settings due to the limited scale and heavy reliance on lexical cues within available benchmarks. Modern benchmarks, inspired by probabilistic causal inference, have attempted to construct causal graphs of events as a robust representation of causal knowledge, where \texttt{CRAB} \citep{romanou2023crab} is one such recent benchmark along this line. In this paper, we introduce \texttt{ACCESS}, a benchmark designed for discovery and reasoning over abstract causal events. Unlike existing resources, \texttt{ACCESS} focuses on causality of everyday life events on the abstraction level. We propose a pipeline for identifying abstractions for event generalizations from \texttt{GLUCOSE} \citep{mostafazadeh-etal-2020-glucose}, a large-scale dataset of implicit commonsense causal knowledge, from which we subsequently extract $1,4$K causal pairs. Our experiments highlight the ongoing challenges of using statistical methods and/or LLMs for automatic abstraction identification and causal discovery in NLP. Nonetheless, we demonstrate that the abstract causal knowledge provided in \texttt{ACCESS} can be leveraged for enhancing QA reasoning performance in LLMs.

Figures

Figures reproduced from arXiv: 2502.08148 by the authors.

Figure 1
Figure 1. Pipeline of abstract causal event discovery. An event is viewed from three hierarchical levels: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. SHD (left) and F1 score (right) of estimated DAGs from statistical structure learning methods. Lower SHD is better. Higher F1 is better. • PC algorithm (Spirtes and Glymour, 1991): A classic approach based on conditional inde￾pendence tests, for which we run two kinds of tests: Chi-squared and G-squared. • DAG-GNN (Yu et al., 2019): DAG structure learning with graph neural networks. • GAE (Ng et al., 2019): This met… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 36 canonical work pages

  1. [1]

    group of participants e.g., a person be playing in the park ̸= a per- son and another person be playing in the park

    single participant vs. group of participants e.g., a person be playing in the park ̸= a per- son and another person be playing in the park

    work page

  2. [2]

    negation e.g., a person be asleep ̸= a person do not sleep

    affirmation vs. negation e.g., a person be asleep ̸= a person do not sleep

    work page

  3. [3]

    future tense e.g., a person go to sleep ̸= a person will go to sleep

    present vs. future tense e.g., a person go to sleep ̸= a person will go to sleep

    work page

  4. [4]

    Acquiring and Modelling Abstract Commonsense Knowledge via Conceptualization

    Knowledge-based news event analysis and forecasting toolkit. In IJCAI, pages 5904–5907. Mutian He, Tianqing Fang, Weiqi Wang, and Yangqiu Song. 2022. Acquiring and modelling abstract com- monsense knowledge via conceptualization. arXiv preprint arXiv:2206.01532. Stefan Heindorf, Yan Scholten, Henning Wachsmuth, Axel-Cyrille Ngonga Ngomo, and Martin Potthast

    work page Pith review arXiv 2022

  5. [5]

    intention or desire e.g., a person do not eat ̸= a person do not want to eat

    work page

  6. [6]

    deduction or possibility e.g., it rain ̸= it may rain

    work page

  7. [7]

    Advances in Neural Information Processing Systems, 33:17943–17954

    On the role of sparsity and dag constraints for learning linear dags. Advances in Neural Information Processing Systems, 33:17943–17954. Ignavier Ng, Shengyu Zhu, Zhitang Chen, and Zhuangyan Fang. 2019. A graph autoencoder ap- proach to causal structure learning. arXiv preprint arXiv:1911.07420. Animesh Nighojkar and John Licato. 2021. Improv- ing paraphr...

    arXiv 2019

  8. [8]

    offers, effort or decision e.g., a person help another person ̸= a person offer to assist an- other person; a person go to the gym ̸= a person decide to go to the gym

    work page

Show all 42 references
  1. [9]

    In Corpus linguistics, volume 2003, page 40

    The timebank corpus. In Corpus linguistics, volume 2003, page 40. Lancaster, UK. James Pustejovsky, Jessica Littman, Roser Saurí, and Marc Verhagen. 2006. Timebank 1.2 documentation. Event London, no. April, pages 6–11. Nils Reimers and Iryna Gurevych. 2020. Making monolingual...

    2003

  2. [10]

    Some sentences describe two actions happening at the same time e.g., a person take something and leave

    multiple actions. Some sentences describe two actions happening at the same time e.g., a person take something and leave. In order to evaluate its meaning, one must select one of them to the key action. The key action is the action that is described by most of other events in ...

  3. [11]

    simple tense

    continuous vs. simple tense. Some sentences describe actions in the continuous state e.g., a person be go home. We ignore the continuous state of the action and consider them equiva- lent to the action described simple tense → a person be go home = a person go home. Step 2 : T...

  4. [14]

    extract conditional independencies from the data distribution to detect edge existence and direc- tion. Meanwhile, score-based methods search for model parameters in the DAG space by optimizing a scoring function (Chickering, 2002; Zheng et al., 2018; Yu et al., 2019; Bello et...

    2023

  5. [15]

    =” to denote semantic similarity and “ ̸=

    is one large-scale annotated corpus that ex- plicitly facilitates causal commonsense knowledge. The dataset captures 10 dimensions of causal expla- nations in story events covering a wider range of en- tities and contexts. GLUCOSE provides rich transla- tions of specific expre...

    2022

  6. [19]

    ability e.g., a person do not eat ̸= a person cannot eat

  7. [22]

    obligation, advice or prohibition e.g.,a person do not eat ̸= a person should not eat

  8. [24]

    In some cases, the object receives an action from the verb refers to a place or location e.g., a person clean a place

    location as object. In some cases, the object receives an action from the verb refers to a place or location e.g., a person clean a place. Here room is considered an (spatial) item be- ing taken action on and similar to any other items such as cup or a table → a person clean a...

  9. [27]

    a causal relation must be temporal, but a tem- poral relation is not always causal

  10. [28]

    the action/state of A directly leads to the ac- tion/state of B i.e., there must be no interme- diate events or if there is one, it should be ex- tremely rare in real-world scenarios

  11. [29]

    Initially, the workers provide non-contextual an- notations based solely on their commonsense under- standing of the abstractions

    an event B would not occur if A did not occur. Initially, the workers provide non-contextual an- notations based solely on their commonsense under- standing of the abstractions. A relation is deemed valid if the annotator can envision a plausible sce- nario in daily life where...

    2019

  12. [31]

    1 |C| |C|X i=1 Aii 2|Ci|

    Self-loop ratio : Proportion of clusters in which the events are either cause or effect of each other. 1 |C| |C|X i=1 Aii 2|Ci|

  13. [32]

    2 |C|2 − |C| |C|−1X i=1 |C|X j=i+1 min(Aij, Aji) max(Aij, Aji)

    Bi-directional ratio: Proportion of cluster pairs that are both cause and effect of one another. 2 |C|2 − |C| |C|−1X i=1 |C|X j=i+1 min(Aij, Aji) max(Aij, Aji)

  14. [33]

    A high value indicates that the object is well matched to its own cluster and poorly matched to neighboring clusters

    Silhouette coefficient (Rousseeuw, 1987): Measure of how similar an instance is to its own cluster (cohesion) compared to other clus- ters (separation). A high value indicates that the object is well matched to its own cluster and poorly matched to neighboring clusters. 1 |D| ...

    1987

  15. [34]

    1 |C| |C|X i=1 2 |Ci|2 − |Ci| X x,y∈Ci,x̸=y Sxy, where Sxy is the similarity score between two events according to Eq

    Homogeneity score: Average pairwise similar- ity of events in a cluster. 1 |C| |C|X i=1 2 |Ci|2 − |Ci| X x,y∈Ci,x̸=y Sxy, where Sxy is the similarity score between two events according to Eq. (1). Table 8 reports the numerical results for the ex- periment on Abstract Event Ide...

    2019

  16. [35]

    Identifying: • Person: John • Action: went, buy • Object: a new collar • Possession: his dog

  17. [36]

    Conversion: a person go to buy another thing for something

  18. [37]

    Original Sentence: John drives near the woman

    Further Conversion: a person buy something to do something ———————————————————- The following is another example. Original Sentence: John drives near the woman

  19. [38]

    Identifying: • Person: John • Action: drives • Object: the woman • Preposition: near

  20. [39]

    Conversion: a person see another person

  21. [40]

    Two women fights each other

    Further Conversion: a person see another person ———————————————————- Now we have a test instance. Please refer to the task instruction and the above examples to do the conversion. The input sentence is: <Input event mention> . Please convert the sentence into a more general ex...

  22. [41]

    In this case, what seems like ‘he made breakfast food’

    Identify the sentence describing ‘another thing which is an ingredient in another thing’: - The relevant sentence from the given text is not explicitly stated but rather implied through analysis. In this case, what seems like ‘he made breakfast food’. In other words - Toaster ...

  23. [42]

    Now find any events occurring after items were used during said activity using index numbers provided against them starting their respective sequence no. This means looking back towards end, start at #5 because there will definitely exist only those indexed things found before...

  24. [1995]

    Psy- chological science, 6(5):292–297

    The construction of situation models in narra- tive comprehension: An event-indexing model. Psy- chological science, 6(5):292–297. A Related Work Theory of causation. Extensive research into theories of causation spans various disciplines (Dalal et al., 2023) such as philosoph...

    2023

  25. [2000]

    MIT press

    Causation, prediction, and search. MIT press. Douglas Steinley. 2004. Properties of the hubert- arable adjusted rand index. Psychological methods, 9(3):386. Yidan Sun, Qin Chao, and Boyang Li. 2023. Event causality is key to computational story understanding. arXiv preprint ar...

    2004 arXiv

  26. [2003]

    Thomas L Griffiths

    10 how does the mind construct and represent stories? Narrative impact: Social and cognitive foundations, page 121. Thomas L Griffiths. 2017. Formalizing prior knowledge in causal induction. The oxford handbook of causal reasoning, pages 115–126. Thomas R Gruber. 1993. A trans...

    2017

  27. [2008]

    and LEIDEN (Traag et al., 2019) algorithms, where LOUVAIN and LEIDEN were proposed for community detection problems. The criteria for selecting these clustering algorithms include: (1) scalability to medium-to-large-sized data, (2) abil- ity to accommodate custom affinity matr...

    2019

  28. [2009]

    Matthew J V owels, Necati Cihan Camgoz, and Richard Bowden

    Information theoretic measures for clusterings comparison: is a correction for chance necessary? In Proceedings of the 26th annual international confer- ence on machine learning, pages 1073–1080. Matthew J V owels, Necati Cihan Camgoz, and Richard Bowden. 2022. D’ya like dags?...

    2022 arXiv

  29. [2013]

    In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1733– 1743

    Why-question answering using intra-and inter- sentential causal relations. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1733– 1743. Judea Pearl. 2009. Causality. Cambridge university press. Judea Pear...

    2009

  30. [2017]

    Oxford hand- book of causal reasoning, pages 1–42

    Mental models and causation. Oxford hand- book of causal reasoning, pages 1–42. Emre Kıcıman, Robert Ness, Amit Sharma, and Chen- hao Tan. 2023. Causal reasoning and large language models: Opening a new frontier for causality. arXiv preprint arXiv:2305.00050. Takeshi Kojima, S...

    2023 arXiv

  31. [2019]

    In Proceed- ings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long Papers)

    Modeling document-level causal structures for event causal relation identification. In Proceed- ings of the 2019 Conference of the North American Chapter of the Association for Computational Lin- guistics: Human Language Technologies, Volume 1 (Long Papers). Roxana Girju. 2003...

    2019

  32. [2020]

    Causenet: Towards a causality graph extracted from the web. In CIKM. ACM. Jena D Hwang, Chandra Bhagavatula, Ronan Le Bras, Jeff Da, Keisuke Sakaguchi, Antoine Bosselut, and Yejin Choi. 2021. (comet-) atomic 2020: On sym- bolic and neural commonsense knowledge graphs. In Proce...

    2021 arXiv

  33. [2022]

    Ad- vances in Neural Information Processing Systems , 35:8226–8239

    Dagma: Learning dags via m-matrices and a log-determinant acyclicity characterization. Ad- vances in Neural Information Processing Systems , 35:8226–8239. Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. 2008. Fast un- folding of communities in l...

    2008 arXiv

  34. [2023]

    arXiv preprint arXiv:2311.04284

    Crab: Assessing the strength of causal rela- tionships between real-world events. arXiv preprint arXiv:2311.04284. Peter J Rousseeuw. 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, ...

    1987 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.