Pith. sign in

REVIEW 6 major objections 5 minor 54 references

GraphRAG-Causal: A novel graph-augmented framework for causal reasoning and annotation in news

T0 review · 6 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read With 20 examples and graph-backed retrieval, a causal news classifier reaches 82.1% F1, close to the supervised baseline.

desk verdict Reasonable GraphRAG-style framework, but the evaluation leaks test sentences into the retrieval graph, so the headline F1 is not evidence. read the letter →

arxiv 2506.11600 v1 pith:42TTACLG submitted 2025-06-13 cs.IR cs.AI

classification cs.IRcs.AI
keywords causalreasoningnewsclassificationgraphretrieval-augmentedgenerationfew-shotlearningknowledgegraphshybridretrievalLLMprompting
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

GraphRAG-Causal tries to show that a large language model can classify causal relationships in news headlines with only 20 labeled examples if the retrieval step supplies structurally matching causal graphs. The pipeline converts annotated sentences into cause-effect-trigger graphs, stores them with embeddings in a graph database, and retrieves the top-k events by a hybrid semantic-plus-structural score. Those retrieved events are inserted into an XML prompt as few-shot demonstrations, and the model returns both a causal label and a tagged sentence. If the claim holds, low-resource causal annotation for news becomes practical without fine-tuning a model, which matters for real-time misinformation and reliability assessment.

What carries the argument

The load-bearing mechanism is the hybrid Cypher query in the graph database. It computes, for every event node with a non-null embedding and text, the cosine similarity between the event embedding and the query embedding, plus a binary structural score that is 1 whenever the event has any cause, effect, or trigger connected to it. The weighted sum $H(e)=\alpha\,\mathrm{sim}(e,q)+\beta S(e)$ selects a small candidate set; the database then collects the texts of linked causes, effects, and triggers for the top candidates, and those texts become the few-shot examples in the XML prompt. The structural term is what distinguishes this from plain vector retrieval: events with parallel causal shape can be retrieved even when their wording is only moderately similar.

What would settle it

Run the pipeline on the Causal News Corpus with an explicit train/test split: build the causal graph and embeddings from the training sentences only, hold out the test sentences, and re-run top-20 retrieval plus prompting on the test set. If the F1 drops far below the reported 82.1% or near majority-class performance, the reported gain depends on retrieving the test sentence rather than learning the causal pattern.

Watch

Extended reading notes

Core claim

The paper's central claim is that a hybrid retrieval score, not prompt engineering alone, carries the few-shot gain. For each event $e$, the score is $H(e)=\alpha\,\mathrm{sim}(e,q)+\beta\,\mathbb{1}\{N_{\text{cause}}(e)+N_{\text{effect}}(e)+N_{\text{trigger}}(e)>0\}$, combining cosine similarity of the query and event embeddings with a binary structural indicator. Events above a threshold are ranked and the top $k$ causal graphs become demonstrations for the LLM. On causal classification over the Causal News Corpus with $k=20$ demonstrations, the framework reports F1 $0.8216$, accuracy $0.7957$, precision $0.7917$, recall $0.8539$, and MCC $0.5856$, close to the fully supervised baseline's F1 $0.8347$; the paper presents this as evidence that graph-augmented few-shot prompting can approach supervised performance.

Load-bearing premise

The framework assumes the retrieval graph contains no test sentences whose labels are being predicted, so the few-shot examples are genuine demonstrations rather than answers; the paper does not document a train/test split, and its own example shows a query sentence retrieved as an identical graph event.

Editorial extensions

If this is right

  • If the 82.1% F1 holds under a clean split, causal news classification no longer requires thousands of labels: twenty gold examples and a prebuilt causal graph suffice.
  • Because the prompt is XML-structured and asks for JSON output, the framework is directly integrable into downstream news-analysis systems without a parsing layer.
  • Increasing k from 5 to 20 raises F1 monotonically for the primary model, implying retrieval knowledge-base size is a controllable lever for accuracy.
  • The claimed domain transfer of generic causal graphs would let the same pipeline move from protest news to finance or health with only a small amount of new annotated data.
  • The approach offers a relatively cheap alternative to fine-tuning an LLM, keeping the base model frozen and using only retrieval plus prompting.

Reading between the lines

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

  • Editorial: the reported F1 is likely sensitive to whether the test sentence itself is stored as a graph node; Section 4.1's example shows an identical query-event pair returning a perfect hybrid score, and the paper does not document a train/test split when building the graph.
  • Editorial: a natural stress test is to rebuild the graph from only the training split and measure the F1 drop on held-out test sentences; until then, the 82.1% figure should be read as an upper-bound estimate.
  • Editorial: the structural score is binary; a count-scaled variant such as $\log(1+N_{\text{cause}})$ might sharpen retrieval among events that are all structurally connected and could be tested against the reported F1.
  • Editorial: the tagging half of the pipeline is self-evaluated by the authors, so the consistency claim for annotations is untested against gold tags; an inter-annotator agreement study against human tagging would settle it.
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

6 major / 5 minor

Summary. The paper proposes GraphRAG-Causal, a three-stage pipeline that converts annotated news sentences from the Causal News Corpus into causal graphs stored in Neo4j, retrieves relevant events through a hybrid score combining embedding cosine similarity with a binary structural score, and feeds the top-k retrieved examples into an LLM via XML-based few-shot prompting for causal classification and cause/effect/trigger tagging. The authors report an F1 of 82.1% at Top K=20 on causal classification, claim consistent gains as the number of retrieved examples increases, and position the approach as a low-data alternative to fine-tuning.

Significance. If the evaluation were trustworthy, the paper would address a useful problem: low-resource causal classification in news without fine-tuning, with explicit cause-effect-trigger tagging. The detailed Cypher query, formal scoring equations, and the use of a public benchmark are strengths. However, the central evidence is compromised: the retrieval store and test set appear to overlap, the structural score is effectively constant for the populated graph, and no non-retrieval baseline or hyperparameter values are reported. As presented, the headline F1 cannot be interpreted as evidence for graph-augmented causal reasoning, and the claimed advantage over normal prompting is not demonstrated.

major comments (6)
  1. [§3.1 and §4] The experimental setup does not separate the retrieval store from the test set. Section 3.1 states that 1030 Causal News Corpus sentences were annotated and inserted into the graph database, while Section 4 states that the test set is also taken from the Causal News Corpus; no disjoint train/test split or deduplication is described. Because the retrieval graph contains the same sentences whose labels are being predicted, the top-k few-shot prompt can include the gold-tagged version of the query itself. Section 4.1 confirms this failure mode: the query sentence is retrieved as Event 1 with a perfect hybrid score of 1.0 with the stated rationale 'identical to the query sentence.' The reported F1 of 0.8216 is therefore not interpretable as evidence for the framework's causal reasoning, since the model may simply copy the supplied answer.
  2. [§3.3, Eqs. (5)-(8) and (14)] The structural score S(e) is a binary indicator that equals 1 whenever the event has at least one causal relation. Since the graph is constructed from 1021 events with a total of 3404 relationships, essentially every candidate event has S(e)=1, so the βS(e) term is a constant added to all hybrid scores. Under that condition the ranking in Eq. (10) is identical to ranking by cosine similarity alone, and the claim that the structural component 'significantly boosts accuracy and consistency' is not supported by the scoring function. The Cypher query in Listing 1 uses the same binary CASE expression. The authors should either use a graded structural score or show that S(e) varies across candidate events and that the ablation with β=0 changes retrieval.
  3. [§3.3 and §4] The hyperparameters that define the method are never reported. Equations (8)-(10) introduce α, β, τ, and top_k, and the Cypher query takes embedding_weight, structure_weight, and similarity_threshold as parameters, but no values for these are given in Section 4. Only top_k is varied. Without these values and without a retrieval ablation (for example, α=1, β=0 versus α=0, β=1 versus the full hybrid), the contribution of the graph-based structural cue cannot be assessed, and the experiments are not reproducible.
  4. [§4, Table 2 and text after Table 2] The text states that 'it can reach 90% accuracy when provided with approximately 50 examples,' but Table 2 reports Top K=50 accuracy of 0.7854 and Top K=20 accuracy of 0.7957. This is an internal contradiction. In addition, the rows of Table 2 are not in monotone order (35, 50, 40 appear after 20), and no run-to-run variance or error bars are reported, so the claimed 'consistent increase in F1' with more examples is not substantiated.
  5. [§4, Tables 2-3] No baseline corresponding to 'normal prompting' is reported. The abstract claims the hybrid retrieval 'significantly boosts accuracy and consistency' compared with normal prompting, but the experimental section only varies top_k for the proposed method and lists the Causal News Corpus supervised baseline. Without a non-retrieval prompt baseline or a random-example few-shot baseline, the improvement claimed for the graph-augmented retrieval component is not demonstrated.
  6. [§4.2 and §5] The paper's only evidence for tagging quality is self-evaluation: Section 4.2 states 'For the tagging accuracy we are self evaluating the tagged sentences,' and Section 5 concedes that 'there is currently no standardized accuracy measure for this component.' Despite this, the abstract and introduction claim improved causal tagging as a contribution. A self-evaluated metric cannot support that claim, and the limitation should be stated in the results section, not only in the conclusion.
minor comments (5)
  1. [§3.1] Section 3.1 states 'manually annotating 1023 sentences,' but the immediately following bullet list says 'Tagged sentences: 1030'; this numeric mismatch should be resolved.
  2. [§4.2, Figures 6-8] The in-text references to Figures 7 and 8 are inaccurate: the text says they illustrate accuracy, precision, recall, and MCC, but Figure 6 shows recall for Maverick, Figure 7 shows accuracy for DeepSeek, and Figure 8 shows MCC for DeepSeek; no precision figure is provided.
  3. [§3.2] Algorithm 1 and the surrounding text refer to 'steps 4 and 5' and 'step 8' without an enumerated step list, making it difficult to follow which component is being described.
  4. [References] The reference list contains duplicates that should be merged: [22] and [43] cite the same survey by Yang et al., and [12] and [45] cite the same paper by Zhao et al.
  5. [§4] The paper does not report the temperature or other sampling settings used for LLM inference, which are needed for reproducibility given the stochasticity of the evaluated models.

Circularity Check

1 steps flagged · score 8.0 of 10

No held-out split: test sentences from Causal News Corpus are stored in the retrieval graph, so the few-shot prompt can contain the exact query sentence with its gold labels; the 82.1% F1 claim reduces to label copying.

  1. self definitional [Section 3.1 (Data Preparation), Section 4 (Experimental Evaluation), Section 4.1 (Table 1)]
    "The first involves acquiring news sentences from Causal News Corpus [25] and start manually annotating 1023 sentences ... These annotated sentences are further processed and converted into causal graphs ... These graphs are then inserted into the graph database ... Test dataset was acquired from Causal News Corpus [25] which contains gold labels with agreement scores. ... This event is identical to the query sentence, with a perfect hybrid score (1.0), embedding similarity (1.0), and structural score (1.0)."

    The retrieval graph is built from gold-annotated Causal News Corpus sentences, and the test set is also drawn from the same corpus with no documented train/test split. Section 4.1 demonstrates the effect directly: the query sentence is returned as Event 1 with perfect similarity because it is identical to an event stored in the graph. Since the few-shot prompt is built from the top-k retrieved events, it can include the very sentence being classified together with its gold cause/effect/trigger annotation. The reported F1 then measures how often the LLM copies the supplied answer rather than how well the framework performs causal reasoning, so the central empirical claim is forced by construction of the retrieval store.

full rationale

The central claim (82.1% F1 at top-20 few-shot retrieval) is not self-contained against the benchmark because the retrieval store and the test set share the same annotated corpus and no disjoint split or deduplication is described. The paper's own Table 1 confirms that an actual query/test sentence is present in the graph as an exact match. This makes the few-shot examples potentially include the gold label for the item being predicted, a reduction of the claimed result to copying. No self-citation chain or ansatz-smuggling issue is present; the circularity is entirely in the evaluation setup. The unreported values of alpha, beta, tau and the irregular top-k sweep are additional correctness risks but are not needed for the circularity finding.

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

All free parameters are unreported or tuned without control. The central assumption of a clean split is contradicted by the paper's own retrieval example. No new entities are postulated beyond standard graph nodes.

free parameters (4)
  • alpha embedding weight = not reported
    Weight on cosine similarity in Eq. 8; chosen by hand or tuned, not justified.
  • beta structural weight = not reported
    Weight on binary structural score in Eq. 8; no sensitivity analysis is provided.
  • threshold tau = not reported
    Filter threshold in Eq. 9; no value or selection procedure is given.
  • top_k = 5, 10, 15, 20, 25, 30, 35, 40, 50 across tables
    Number of few-shot examples; swept over many values with inconsistent ordering, suggesting post hoc selection on the test set.
assumptions (5)
  • domain assumption The retrieval graph is disjoint from the evaluation test set.
    Required for unbiased evaluation; contradicted by Section 4.1 exact match example and no split described.
  • domain assumption Hand-annotated causal graphs (causes, effects, triggers) are accurate and complete.
    Section 3.1 manual annotation; no inter-annotator agreement or validation is reported.
  • domain assumption all-MiniLM-L6-v2 cosine similarity is a valid semantic relevance signal for causal events.
    Used in Eq. 7 without task-specific evaluation.
  • ad hoc to paper Binary structural score S(e) (presence of any relation) improves retrieval.
    Introduced in Eq. 5; no ablation establishes its contribution.
  • domain assumption Causal News Corpus gold labels are ground truth.
    Test labels taken from CNC [25] with no discussion of label noise beyond crowd agreement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GraphRAG-Causal: A novel graph-augmented framework for causal reasoning and annotation in news." pith.science (2026). https://pith.science/paper/42TTACLG

@misc{pith2026250611600,
  author       = {Pith},
  title        = {Pith review of: GraphRAG-Causal: A novel graph-augmented framework for causal reasoning and annotation in news},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42TTACLG}},
  note         = {Machine review of arXiv:2506.11600}
}
read the original abstract

GraphRAG-Causal introduces an innovative framework that combines graph-based retrieval with large language models to enhance causal reasoning in news analysis. Traditional NLP approaches often struggle with identifying complex, implicit causal links, especially in low-data scenarios. Our approach addresses these challenges by transforming annotated news headlines into structured causal knowledge graphs. It then employs a hybrid retrieval system that merges semantic embeddings with graph-based structural cues leveraging Neo4j to accurately match and retrieve relevant events. The framework is built on a three-stage pipeline: First, during Data Preparation, news sentences are meticulously annotated and converted into causal graphs capturing cause, effect, and trigger relationships. Next, the Graph Retrieval stage stores these graphs along with their embeddings in a Neo4j database and utilizes hybrid Cypher queries to efficiently identify events that share both semantic and structural similarities with a given query. Finally, the LLM Inference stage utilizes these retrieved causal graphs in a few-shot learning setup with XML-based prompting, enabling robust classification and tagging of causal relationships. Experimental evaluations demonstrate that GraphRAG-Causal achieves an impressive F1-score of 82.1% on causal classification using just 20 few-shot examples. This approach significantly boosts accuracy and consistency, making it highly suitable for real-time applications in news reliability assessment, misinformation detection, and policy analysis.

Figures

Figures reproduced from arXiv: 2506.11600 by the authors.

Figure 1
Figure 1. Overview of the proposed approach, illustrating a three-stage pipeline: (1) Data Preparation (data annotation [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. illustrates the overall workflow of this approach. The query sentence is first encoded into an embedding and analyzed for structural components. The resulting scores are then merged, and events are ranked by their final hybrid score. Input Query Sentence Embedding Generation Structural Analysis Hybrid Score Calculation (Embedding + Structural) Event Retrieval [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Hybrid Scores for the Retrieved Similar Events. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: F1 Score Growth for DeepSeek-R1-Distill-LLaMA-70B [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: F1 Score Growth for LLaMA 4 Maverick 17B Instruct 128B [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Recall Score Comparison for LLaMA 4 Maverick 17B Instruct 128B across Iterations [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Accuracy Trend for DeepSeek-R1-Distill-LLaMA-70B [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: MCC Trend for DeepSeek-R1-Distill-LLaMA-70B [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 37 canonical work pages

  1. [1]

    MLModeler5@ Causal News Corpus 2023: Us- ing RoBERTa for Casual Event Classification

    Bhatia, Amrita, Thomas, Ananya, Jain, Nitansh, and Bedi, Jatin. "MLModeler5@ Causal News Corpus 2023: Us- ing RoBERTa for Casual Event Classification." Proceedings of the 6th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text, pp. 34–37, 2023

  2. [2]

    Investigating Causal Reasoning in Large Language Models

    Rawal, Atul, Raglin, Adrienne, Wang, Qianlong, and Tang, Ziying. "Investigating Causal Reasoning in Large Language Models." NeurIPS 2024 Workshop on Causality and Large Models (CaLM) , 2024. URL: https: //openreview.net/pdf?id=EGWrfirmIM

  3. [3]

    The Role of Causality in Explainable Artificial Intelli- gence

    Carloni, Gianluca, Berti, Andrea, and Colantonio, Sara. "The Role of Causality in Explainable Artificial Intelli- gence." arXiv preprint arXiv:2309.09901, 2023. URL: https://arxiv.org/abs/2309.09901

  4. [4]

    Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond

    Feder, Amir, Keith, Katherine A., Manzoor, Emaad, Pryzant, Reid, Sridhar, Dhanya, Wood-Doughty, Zach, Eisenstein, Jacob, Grimmer, Justin, Reichart, Roi, Roberts, Margaret E., Stewart, Brandon M., Veitch, Victor, and Yang, Diyi. "Causal Inference in Natural Language Processing: Estimation, Prediction, Interpretation and Beyond." arXiv preprint arXiv:2109.0...

  5. [5]

    CausalNLP Tutorial: An Introduction to Causality for Natural Language Processing

    Jin, Zhijing, Feder, Amir, and Zhang, Kun. "CausalNLP Tutorial: An Introduction to Causality for Natural Language Processing." In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing: Tutorial Abstracts , pp. 17–22. Association for Computational Linguistics, 2022. URL: https: //aclanthology.org/2022.emnlp-tutorials.4.pdf

  6. [6]

    Using Natural Language Processing to Extract Health-Related Causality from Twitter Messages

    Doan, Son, Yang, Elly W., Tilak, Sameer, and Torii, Manabu. "Using Natural Language Processing to Extract Health-Related Causality from Twitter Messages." In Proceedings of the 2018 IEEE International Conference on Healthcare Informatics Workshop (ICHI-W), pp. 47–48. IEEE, 2018. DOI: 10.1109/ICHI-W.2018.00031

  7. [7]

    HeadlineCause: A Dataset of News Headlines for Detecting Causalities

    Gusev, Ilya, and Tikhonov, Alexey. "HeadlineCause: A Dataset of News Headlines for Detecting Causalities." arXiv preprint arXiv:2108.12626, 2021. URL: https://arxiv.org/abs/2108.12626

  8. [8]

    Causality for Trustworthy Artificial Intelligence: Status, Challenges, and Opportu- nities

    Asiaee, Amir, and Liu, Huan. "Causality for Trustworthy Artificial Intelligence: Status, Challenges, and Opportu- nities." Communications of the ACM, vol. 66, no. 12, pp. 44–53, Dec. 2023. DOI: 10.1145/3665494

Show all 54 references
  1. [9]

    ARGUABLY @ Causal News Corpus 2022: Contextually Augmented Language Models for Event Causality Identification

    Kohli, Guneet Singh, Kaur, Prabsimran, and Bedi, Jatin. "ARGUABLY @ Causal News Corpus 2022: Contextually Augmented Language Models for Event Causality Identification." Proceedings of the 5th Workshop on Challenges and Applications of Automated Extraction of Socio-political Ev...

  2. [10]

    Retrieval-Augmented Generation for Large Language Models: A Survey

    Gao, Yunfan, Xiong, Yun, Gao, Xinyu, Jia, Kangxiang, Pan, Jinliu, Bi, Yuxi, Dai, Yi, Sun, Jiawei, and Wang, Haofen. "Retrieval-Augmented Generation for Large Language Models: A Survey." arXiv preprint arXiv:2312.10997, December 2023. URL: https://arxiv.org/abs/2312.10997

  3. [11]

    Retrieval-Augmented Generation with Graphs (GraphRAG)

    Han, Haoyu, Wang, Yu, Shomer, Harry, Guo, Kai, Ding, Jiayuan, Lei, Yongjia, Halappanavar, Mahantesh, Rossi, Ryan A., Mukherjee, Subhabrata, Tang, Xianfeng, He, Qi, Hua, Zhigang, Long, Bo, Zhao, Tong, Shah, Neil, Javari, Amin, Xia, Yinglong, and Tang, Jiliang. "Retrieval-Augmen...

  4. [12]

    Causal-CoG: A Causal-Effect Look at Context Generation for Boosting Multi-modal Language Models

    Zhao, Shitian, Li, Zhuowan, Lu, Yadong, Yuille, Alan, and Wang, Yan. "Causal-CoG: A Causal-Effect Look at Context Generation for Boosting Multi-modal Language Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. URL: https://...

  5. [13]

    Causal Inference and Natural Language Processing

    Chen, Wenqing, and Chu, Zhixuan. "Causal Inference and Natural Language Processing." In Machine Learning for Causal Inference, edited by Sheng Li and Zhixuan Chu, pp. 189–206. Springer International Publishing, 2023. DOI: 10.1007/978-3-031-35051-1_9

  6. [14]

    CSECU-DSG @ Causal News Corpus 2022: Fusion of RoBERTa Transformers Variants for Causal Event Classification

    Aziz, Abdul, Hossain, Md. Akram, and Chy, Abu Nowshed. "CSECU-DSG @ Causal News Corpus 2022: Fusion of RoBERTa Transformers Variants for Causal Event Classification." In Proceedings of the 5th Workshop on Challenges and Applications of Automated Extraction of Socio-political E...

  7. [15]

    Causal Graphs Meet Thoughts: Enhancing Complex Reasoning in Graph-Augmented LLMs

    Luo, Hang, Zhang, Jian, and Li, Chujun. "Causal Graphs Meet Thoughts: Enhancing Complex Reasoning in Graph-Augmented LLMs." arXiv preprint arXiv:2501.14892, January 2025. URL: https://arxiv.org/abs/ 2501.14892

  8. [16]

    NLP4ITF @ Causal News Corpus 2022: Leveraging Linguistic Infor- mation for Event Causality Classification

    Krumbiegel, Theresa, and Decher, Sophie. "NLP4ITF @ Causal News Corpus 2022: Leveraging Linguistic Infor- mation for Event Causality Classification." In Proceedings of the 5th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (...

  9. [17]

    Causal graph extraction from news: a comparative study of time-series causality learning techniques

    Maisonnave, Mariano, Delbianco, Fernando, Tohmé, Fernando, Milios, Evangelos, and Maguitman, Ana G. "Causal graph extraction from news: a comparative study of time-series causality learning techniques." PeerJ Computer Science, vol. 8, article e1066, Aug. 2022. DOI: 10.7717/pee...

  10. [18]

    Causality: Models, Reasoning, and Inference

    Pearl, Judea. Causality: Models, Reasoning, and Inference . 2nd ed., Cambridge University Press, 2009. DOI: 10.1017/CBO9780511803161

  11. [19]

    Iden- tifying Predictive Causal Factors from News Streams

    Balashankar, Ananth, Chakraborty, Sunandan, Fraiberger, Samuel, and Subramanian, Lakshminarayanan. "Iden- tifying Predictive Causal Factors from News Streams." In Proceedings of the 2019 Conference on Empiri- cal Methods in Natural Language Processing and the 9th International...

  12. [20]

    Causal Understanding of Fake News Dissemination on Social Media

    Cheng, Lu, Guo, Ruocheng, Shu, Kai, and Liu, Huan. "Causal Understanding of Fake News Dissemination on Social Media." In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2021), pp. 148–157. Association for Computing Machinery, 2021. DOI...

  13. [21]

    NoisyAnnot@ Causal News Corpus 2022: Causality Detection using Multiple Annotation Decisions

    Nguyen, Quynh Anh, and Mitra, Arka. "NoisyAnnot@ Causal News Corpus 2022: Causality Detection using Multiple Annotation Decisions." Proceedings of the 5th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (CASE) , pages 79–84, ...

  14. [22]

    A survey on extraction of causal relations from nat- ural language text

    Yang, Jie, Han, Soyeon Caren, and Poon, Josiah. "A survey on extraction of causal relations from nat- ural language text." Knowledge and Information Systems , vol. 64, pp. 1161–1186, 2022. DOI: 10.1007/ s10115-022-01665-w

  15. [23]

    Text to causal knowledge graph: A framework to synthesize knowledge from unstructured business texts into causal graphs

    Gopalakrishnan, Seethalakshmi, Chen, Victor Zitian, Dou, Wenwen, Hahn-Powell, Gus, Nedunuri, Sreekar, and Zadrozny, Wlodek. "Text to causal knowledge graph: A framework to synthesize knowledge from unstructured business texts into causal graphs." Information, vol. 14, no. 7, p...

  16. [24]

    Causal Representation Learning with Generative Artificial Intelligence: Application to Texts as Treatments

    Imai, Kosuke, and Nakamura, Kentaro. "Causal Representation Learning with Generative Artificial Intelligence: Application to Texts as Treatments." arXiv preprint arXiv:2410.00903, 2024. URL: https://arxiv.org/abs/ 2410.00903

  17. [25]

    The causal news corpus: Annotating causal relations in event sentences from news

    Tan, Fiona Anting, Hürriyeto ˘glu, Ali, Caselli, Tommaso, Oostdijk, Nelleke, Nomoto, Tadashi, Hettiarachchi, Hansi, Ameer, Iqra, Uca, Onur, Liza, Farhana Ferdousi, and Hu, Tiancheng. "The causal news corpus: Annotating causal relations in event sentences from news." arXiv prep...

  18. [26]

    Causality for Machine Learning

    Schölkopf, Bernhard. "Causality for Machine Learning." In Probabilistic and Causal Inference: The Works of Judea Pearl, edited by Hector Geffner, Rina Dechter, and Joseph Y . Halpern, 765–804. New York, NY , USA: Association for Computing Machinery, 2022. DOI: 10.1145/3501714.3501755

  19. [27]

    Constructing and interpreting causal knowledge graphs from news

    Tan, Fiona Anting, Paul, Debdeep, Yamaura, Sahim, Koji, Miura, and Ng, See-Kiong. "Constructing and interpreting causal knowledge graphs from news." Proceedings of the AAAI Symposium Series, vol. 1, no. 1, pp. 52–59, 2023

  20. [28]

    Causal reasoning and large language models: Opening a new frontier for causality

    Kıciman, Emre, Ness, Robert, Sharma, Amit, and Tan, Chenhao. "Causal reasoning and large language models: Opening a new frontier for causality." arXiv preprint arXiv:2305.00050, 2023

  21. [29]

    Causal Reasoning in Large Language Models using Causal Graph Retrieval Augmented Generation

    Samarajeewa, Chamod, De Silva, Daswin, Osipov, Evgeny, Alahakoon, Damminda, and Manic, Milos. "Causal Reasoning in Large Language Models using Causal Graph Retrieval Augmented Generation." 2024 16th Interna- tional Conference on Human System Interaction (HSI), pp. 1–6, IEEE, 2024

  22. [30]

    Groq - Accelerating AI Workloads

    Groq. "Groq - Accelerating AI Workloads." Available: https://groq.com/. Accessed: Nov. 17, 2024

  23. [31]

    A Survey of Learning Causality with Data: Problems and Methods

    Guo, Ruocheng, Cheng, Lu, Li, Jundong, Hahn, P. Richard, and Liu, Huan. "A Survey of Learning Causality with Data: Problems and Methods." ACM Computing Surveys, vol. 53, no. 4, article 75, pp. 1–37, July 2021. DOI: 10.1145/3397269

  24. [32]

    Learning Causality for News Events Prediction

    Radinsky, Kira, Davidovich, Sagie, and Markovitch, Shaul. "Learning Causality for News Events Prediction." In Proceedings of the 21st International Conference on World Wide Web (WWW 2012) , pp. 909–918. Association for Computing Machinery, 2012. DOI: 10.1145/2187836.2187958

  25. [33]

    Investigating causal understanding in LLMs

    Hobbhahn, Marius, Lieberum, Tom, and Seiler, David. "Investigating causal understanding in LLMs." NeurIPS ML Safety Workshop, 2022. URL: https://openreview.net/forum?id=FQI5KxgFRc

  26. [34]

    Causal Intervention and Counterfactual Reasoning for Multi-modal Fake News Detection

    Chen, Ziwei, Hu, Linmei, Li, Weixin, Shao, Yingxia, and Nie, Liqiang. "Causal Intervention and Counterfactual Reasoning for Multi-modal Fake News Detection." In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pp...

  27. [35]

    The explanation of a complex problem: A content analysis of causality in cancer news

    Peng, Wei, de Tuya, Gabriel Alexander, Eduardo, Andrea Alexandra, Vishny, Jessica Allison, and Huang, Qian. "The explanation of a complex problem: A content analysis of causality in cancer news." Public Understanding of Science, vol. 30, no. 7, pp. 857–872, Oct. 2021. DOI: 10....

  28. [36]

    Chain-of-thought prompting elicits reasoning in large language models

    Wei, Jason, Wang, Xuezhi, Schuurmans, Dale, Bosma, Maarten, Xia, Fei, Chi, Ed, Le, Quoc V ., Zhou, Denny, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in Neural Information Processing Systems, vol. 35, pp. 24824–24837, 2022

  29. [37]

    Causal parrots: Large language models may talk causality but are not causal

    Zeˇcevi´c, Matej, Willig, Moritz, Dhami, Devendra Singh, and Kersting, Kristian. "Causal parrots: Large language models may talk causality but are not causal." arXiv preprint arXiv:2308.13067, 2023

  30. [38]

    Prompt-based vs. Fine-tuned LLMs Toward Causal Graph Verification

    Susanti, Yuni, and Holsmoelle, Nina. "Prompt-based vs. Fine-tuned LLMs Toward Causal Graph Verification." arXiv preprint arXiv:2406.16899, 2024

  31. [39]

    Cause and Effect: Can Large Language Models Truly Understand Causality?

    Ashwani, Swagata, Hegde, Kshiteesh, Reddy Mannuru, Nishith, Singh Sengar, Dushyant, Jindal, Mayank, Chaitanya Rao Kathala, Krishna, Banga, Dishant, Jain, Vinija, and Chadha, Aman. "Cause and Effect: Can Large Language Models Truly Understand Causality?" Proceedings of the AAAI...

  32. [40]

    Evaluation Methods and Measures for Causal Learning Algorithms

    Cheng, Lu, Guo, Ruocheng, Moraffah, Raha, Sheth, Paras, Candan, K. Selçuk, and Liu, Huan. "Evaluation Methods and Measures for Causal Learning Algorithms." IEEE Transactions on Artificial Intelligence, vol. 3, no. 6, pp. 924–943, 2022. DOI: 10.1109/TAI.2022.3150264

  33. [41]

    Pairwise Causality Guided Transformers for Event Sequences

    Shou, Xiao, Bhattacharjya, Debarun, Gao, Tian, Subramanian, Dharmashankar, Hassanzadeh, Oktie, and Bennett, Kristin P. "Pairwise Causality Guided Transformers for Event Sequences." Advances in Neural Information Processing Systems, vol. 36, pp. 46520–46533, 2023

  34. [42]

    all-MiniLM-L6-v2 - Sentence Transformers

    Hugging Face. "all-MiniLM-L6-v2 - Sentence Transformers." Available: https://huggingface.co/ sentence-transformers/all-MiniLM-L6-v2 . Accessed: Dec. 17, 2024

  35. [43]

    A survey on extraction of causal relations from natural language text

    Yang, Jie, Han, Soyeon Caren, and Poon, Josiah. "A survey on extraction of causal relations from natural language text." Knowledge and Information Systems, vol. 64, no. 5, pp. 1161–1186, 2022

  36. [44]

    Causal knowledge extraction from long text maintenance documents

    Hershowitz, Brad, Hodkiewicz, Melinda, Bikaun, Tyler, Stewart, Michael, and Liu, Wei. "Causal knowledge extraction from long text maintenance documents." Computers in Industry, vol. 161, pp. 104110, 2024

  37. [45]

    Causal-CoG: A Causal-Effect Look at Context Generation for Boosting Multi-modal Language Models

    Zhao, Shitian, Li, Zhuowan, Lu, Yadong, Yuille, Alan, and Wang, Yan. "Causal-CoG: A Causal-Effect Look at Context Generation for Boosting Multi-modal Language Models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 13342–13351, 2024

  38. [46]

    CHEER: Centrality-aware high-order event reasoning network for document-level event causality identification

    Chen, Meiqi, Cao, Yixin, Zhang, Yan, and Liu, Zhiwei. "CHEER: Centrality-aware high-order event reasoning network for document-level event causality identification." Proceedings of The 61st Annual Meeting of the Association for Computational Linguistics, ACM, 2023

  39. [47]

    Root Cause Analysis in Microservice Using Neural Granger Causal Discovery

    Lin, Cheng-Ming, Chang, Ching, Wang, Wei-Yao, Wang, Kuang-Da, and Peng, Wen-Chih. "Root Cause Analysis in Microservice Using Neural Granger Causal Discovery." Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 1, pp. 206–213, 2024. DOI: 10.1609/aaai.v3...

  40. [48]

    Financial Causal Sentence Recognition Based on BERT-CNN Text Classification

    Wan, Chang-Xuan, and Li, Bo. "Financial Causal Sentence Recognition Based on BERT-CNN Text Classification." The Journal of Supercomputing, vol. 78, no. 5, pp. 6503–6527, 2022. DOI: 10.1007/s11227-021-04097-5

  41. [49]

    The Financial Document Causality Detection Shared Task (FinCausal 2023)

    Moreno-Sandoval, Antonio, Porta-Zamorano, Jordi, Carbajo-Coronado, Blanca, Samy, Doaa, Mariko, Dominique, and El-Haj, Mahmoud. "The Financial Document Causality Detection Shared Task (FinCausal 2023)." 2023 IEEE International Conference on Big Data (BigData), pp. 2855–2860, 2023

  42. [50]

    Event Causality Extraction via Implicit Cause-Effect Interactions

    Liu, Jintao, Zhang, Zequn, Wei, Kaiwen, Guo, Zhi, Sun, Xian, Jin, Li, and Li, Xiaoyu. "Event Causality Extraction via Implicit Cause-Effect Interactions." Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Association for Computational Ling...

  43. [51]

    End-to-end multi-granulation causality extraction model

    Wu, Miao, Zhang, Qinghua, Wu, Chengying, and Wang, Guoyin. "End-to-end multi-granulation causality extraction model." Digital Communications and Networks, 2023. DOI: 10.1016/j.dcan.2023.02.005

  44. [52]

    Neo4j Resources

    Neo4j. "Neo4j Resources." Available: https://neo4j.com/resources/. Accessed: Dec. 10, 2024. 18

  45. [2022]

    URL: https://aclanthology.org/2022.case-1.20.pdf

  46. [2023]

    17 A PREPRINT - AUGUST 28, 2025

    URL: https://aclanthology.org/2023.acl-long.37/. 17 A PREPRINT - AUGUST 28, 2025

Pith tools

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