Pith. sign in

REVIEW 4 major objections 5 minor 50 references

Retrieval Augmented Generation based Large Language Models for Causality Mining

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read This paper claims that dynamically retrieving few-shot causal examples by causal-connective match and semantic similarity improves LLM performance on causality detection and extraction, beating static prompting and a supervised…

desk verdict A sensible incremental RAG idea for causality mining that is oversold in the abstract; the detection pipeline needs specification before the results are reproducible. read the letter →

arxiv 2505.23944 v1 pith:3AJBO7H4 submitted 2025-05-29 cs.CL

classification cs.CL
keywords causalitydetectionextractionretrieval-augmentedgenerationin-contextlearningfew-shotpromptinglargelanguagemodelscausalconnectives
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 attempts to establish that large language models do causality detection and extraction better when the few-shot examples in the prompt are chosen dynamically by retrieval, rather than fixed in advance. The proposed Pattern RAG scheme retrieves examples that share the input sentence's causal connective, and kNN+Pattern RAG adds the ten semantically nearest examples. On three datasets (SemEval, the ADE drug-effect corpus, and a multi-cause-effect dataset) and five LLMs, the dynamic prompts reportedly outperform zero-shot, random few-shot, and semantic-only kNN baselines, and at their best match or exceed a supervised transformer on extraction accuracy. The significance is a low-cost route to causal knowledge extraction that does not require fine-tuning or large domain-specific training sets.

What carries the argument

The load-bearing object is the few-shot example repository: 2,365 causal sentences with cause and effect spans tagged and indexed by causal connectives, the expressions like 'caused by' or 'lead to' that link cause to effect, which were identified by prompting GPT-3.5-turbo. Pattern RAG retrieves by matching the input's causal connective to these indexes at a 90 percent similarity threshold; kNN+Pattern RAG concatenates the pattern matches with ten nearest neighbors under text-embedding-ada-002 embeddings. The machinery does two jobs: it supplies the LLM with examples that use the same syntactic pattern for extraction, and with semantically similar sentences for detection, testing the hypothesis that both forms of relevance matter more than example count.

What would settle it

On the same three detection test sets, run kNN+Pattern RAG with the pattern-retrieved examples replaced by an equal number of randomly sampled repository examples while keeping the kNN examples and prompt format fixed; if F1 does not fall, the causal-connective index is not the active ingredient.

Watch

Extended reading notes

Core claim

The paper's central claim is that retrieval-augmented prompting with causally relevant examples is the decisive ingredient for getting LLMs to find cause-effect structure in text. Pattern RAG matches the input's causal connective against an offline repository indexed by causally tagged examples and retrieves those above a 90 percent similarity; kNN+Pattern RAG combines those with ten embedding-nearest neighbors. The paper reports that the combined scheme gives the strongest detection results for the best-performing model, and that pattern-based retrieval in particular improves cause-effect phrase extraction, lifting F1 by up to 4 points over the best baseline on the multi-cause-effect dataset and matching or beating the supervised DEPBERT model on SemEval extraction accuracy. Detection F1 gains over the strongest baseline are reported as 2, 7, and 1 percentage points on the three datasets.

Load-bearing premise

The load-bearing premise is that every input sentence, including non-causal ones, has a causal connective that can be identified reliably enough to retrieve matching examples; the paper never specifies how that connective is obtained for a sentence before retrieval.

Editorial extensions

If this is right

  • Causality detection and extraction can be improved without fine-tuning, a concrete advantage because supervised causality models are known to lose accuracy across domains.
  • A repository of a few thousand tagged causal sentences is enough to produce these gains, so large labeled training corpora are not a prerequisite.
  • Relevance of the retrieved examples matters more than their number: RAG-based prompts stay ahead of random few-shot prompting even when the random prompt is given more examples.
  • Pattern-based retrieval by causal connective is sufficient to help cause-effect phrase extraction, and adding semantic neighbors helps most in detection tasks.
  • At its best, the method matches or exceeds a supervised transformer baseline on a standard causality extraction benchmark.

Reading between the lines

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

  • Editorial inference: if the pattern-matching step is what drives extraction gains, the same retrieval design could transfer to other marked linguistic relations such as temporal, contrastive, or conditional relations by re-indexing the repository with their connectives.
  • Editorial inference: because the paper only handles within-sentence causality, a natural extension would index discourse-level connectives across sentence pairs; the reported detection gains would need a defined way to assign a connective to non-causal inputs.
  • Editorial inference: the exact-match evaluation penalizes phrase-boundary differences such as a missing article, so part of the observed advantage on multi-word extraction may reflect evaluation strictness; a span-matching metric would reveal the remaining true gain.
  • Editorial inference: a cheap ablation, adding random repository examples on top of kNN rather than pattern-matched ones, would isolate whether causal-connective indexing or simply more examples causes the improvement.
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

4 major / 5 minor

Summary. The paper proposes two retrieval-augmented prompting strategies, Pattern RAG and kNN+Pattern RAG, for causality detection and cause/effect extraction with LLMs. A Fewshot Example DB is built offline from causal training sentences of SemEval, ADE, and Li et al., indexed by causal connectives extracted with GPT-3.5-turbo. Pattern RAG retrieves examples whose connective is more than 90% similar to the input's connective; kNN RAG retrieves the 10 nearest sentences by text-embedding-ada-002; kNN+Pattern RAG concatenates the retrieved examples from both methods. The paper reports detection and extraction results over three datasets and five LLMs, comparing against zero-shot, random few-shot, kNN RAG, and supervised baselines. The abstract and conclusion claim that the proposed RAG-based dynamic prompting significantly outperforms static prompting schemes on both causality detection and extraction tasks. The code is released on GitHub.

Significance. If the method is made fully reproducible and the performance claims are appropriately qualified, the paper would be a useful contribution: it introduces a lightweight, training-free retrieval signal (causal connectives) that is intuitively aligned with the cause/effect annotation task, evaluates across multiple LLMs and datasets, and studies how few-shot count interacts with selection quality. The kNN+Pattern combination is a simple but sensible way to use both semantic and syntactic cues. However, the current evidence does not support the unqualified superiority claim: several table cells show ties or losses, and the detection protocol for Pattern RAG is underspecified. The paper also demonstrates awareness of limitations in Section 9, but these do not cover the reproducibility issues identified below.

major comments (4)
  1. [§3.2 and §4] Pattern RAG's retrieval rule is undefined for the causality-detection task on non-causal inputs. Detection inputs are defined only as sentences labeled 1 or 0 (§4), but §3.2 says Pattern RAG selects examples by matching "the causal connective of the input sentence" against the Fewshot Example DB. No procedure is given for obtaining a causal connective for a 0-labeled sentence, nor for what to do when no connective is found. Since Table 3 reports Pattern RAG and kNN+Pattern RAG on SemEval (300 non-causal), ADE (2,000 non-causal), and Li et al. (595 non-causal) task-1 splits (Table 9), the detection half of the central claim cannot be reproduced from the paper. The paper should specify the fallback retrieval for non-causal inputs and, if no fallback is used, state explicitly that the method only applies to causal inputs.
  2. [§5, Tables 3 and 4] The abstract and conclusion claim that the proposed RAG methods significantly outperform baselines "for both causality detection and causality extraction tasks" is not supported by the reported numbers. In Table 4, SemEval extraction for GPT-3.5-turbo gives 0.91 accuracy for both kNN RAG and kNN+Pattern RAG, and Llama3-8b yields identical 0.83 accuracy for kNN RAG, Pattern RAG, and kNN+Pattern RAG. In Table 3, for GPT-4o on SemEval detection, Pattern RAG F1 (0.86) is lower than kNN RAG F1 (0.88); for Gemma2-9b-it on SemEval detection, Pattern RAG F1 (0.76) is lower than kNN RAG F1 (0.80). These observations are inconsistent with a blanket claim of superiority; the authors should revise the claim and identify the settings and LLMs where each proposed method is actually beneficial.
  3. [§3.2, §4.1] The primary comparison is confounded by the number of in-context examples. Random fewshot and kNN RAG use 10 examples, while kNN+Pattern RAG concatenates 10 kNN examples with up to 10 Pattern RAG examples, yielding up to 20 examples. In Tables 3 and 4, the gains of kNN+Pattern RAG over kNN RAG could therefore be due to more in-context examples rather than to the retrieval strategy. Section 6 studies example-count curves, but the main tables do not include an equal-count control (e.g., kNN RAG with 20 examples or kNN+Pattern RAG with 5+5 examples). The authors should add such a control or explicitly restrict kNN+Pattern RAG to match the baseline count.
  4. [§3.2] The similarity criterion and stochastic selection make Pattern RAG irreproducible as specified. The paper says "more than 90 percent similar" without defining the metric on strings (edit distance, token overlap, embedding cosine?). It also says the repository keeps "up to 10 random examples per unique causal connective" and, in §3.2, "randomly choose 10 of them" when more matches exist. Without a seeded sampling scheme or multiple-run reporting, two implementations can produce different prompts and different Table 3 and Table 4 numbers. Define the similarity measure, and either make retrieval deterministic (e.g., first-k by a tie-breaker) or report mean and variance over seeds.
minor comments (5)
  1. [§3.1, Figure 1] The reference to Figure 1 appears as "Figure ??" in the text, and the figure itself is not included in the manuscript.
  2. [§4, Datasets] For the ADE detection task, the 2,000 non-causal sentences are drawn from the Li et al. training data rather than from ADE; this makes the ADE detection benchmark cross-domain, and the choice should be justified or discussed.
  3. [§4, Evaluation] For SemEval and ADE extraction, a prediction is considered correct if the ground-truth word appears in a larger predicted phrase; this is a lenient containment metric rather than exact match and should be stated in the main text, especially when comparing with DEPBERT's reported scores.
  4. [Table 6] The table caption refers to blue and red colors for cause and effect labels, but the table as printed uses ✓/× marks; the formatting should be clarified so the reader can map symbols and colors to predictions.
  5. [§3.1] Since GPT-3.5-turbo is used both to build the causal-connective index and as one of the evaluated LLMs, the authors should note that Pattern RAG may be especially well tuned to that model's connective annotations, and discuss whether the conclusions transfer to other LLMs.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical comparison with no fitted equation or self-citation chain that reduces the central claim to its inputs.

full rationale

The central claim ('our RAG-based dynamic prompting over other static prompting schemes' is superior) is supported by direct experiments on held-out test splits. The Fewshot Example DB is built offline from training splits, and the paper explicitly states that test causal sentences are absent from it (Section A, Table 9: 'we select 249 causal sentences from the test set that are absent in our Fewshot Example DB'). No parameter is fitted to the test predictions; the 90% similarity threshold, the 10-example cap, and the k in kNN are fixed hyperparameters, not fitted outputs. The repository indexing uses GPT-3.5-turbo to extract causal connectives, and GPT-3.5-turbo is also one of the evaluated models, but the retrieval key is not the predicted cause/effect label, so this is not a definitional collapse. The main weakness is a reproducibility gap rather than circularity: Section 3.2 defines Pattern RAG by matching 'the causal connective of the input sentence' against the repository, but Section 4 defines detection inputs only as sentences labeled 1 or 0 and never specifies how a causal connective is obtained for a non-causal sentence (nor is the 90% similarity measure defined). This makes the Table 3 detection results for Pattern RAG and kNN+Pattern RAG hard to reproduce from the text, but it does not make the claim equivalent to its inputs by construction. There are no load-bearing self-citations or imported uniqueness theorems.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No free parameters are fitted to the test data in a numeric sense, but four hyperparameters are chosen by hand and affect the central comparison. The Fewshot Example DB is a constructed artifact, not a postulated theoretical entity. The axioms are domain assumptions about retrieval signals and repository coverage.

free parameters (4)
  • k in kNN retrieval = 10
    Chosen by hand in Section 3.2; no ablation or sensitivity analysis for k.
  • Pattern RAG connective similarity threshold = >90%
    Chosen by hand in Section 3.2; the paper does not study the effect of this threshold.
  • Examples per causal connective cap = 10
    Set for memory efficiency in Section 3.1; the authors acknowledge in Section 6 that this cap limits Pattern RAG when more examples are requested.
  • Combined kNN+Pattern example count = 20
    Concatenation of two 10-example sets in Section 3.2; no ablation isolating the contribution of the doubled count.
assumptions (3)
  • domain assumption Causal connectives are a sufficient retrieval signal for useful few-shot examples.
    Section 3.2 states this as the Pattern RAG hypothesis; Section 3.1 only spot-checks connective quality manually.
  • domain assumption Sentence embeddings from text-embedding-ada-002 capture semantic similarity relevant to causality.
    Section 3.2 uses this embedding for kNN retrieval without an embedding ablation.
  • domain assumption A small curated repository of 2,365 sentences built from training splits covers test-domain causal patterns.
    Section 3.1 states repository desiderata but provides no coverage analysis over the three test distributions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Retrieval Augmented Generation based Large Language Models for Causality Mining." pith.science (2026). https://pith.science/paper/3AJBO7H4

@misc{pith2026250523944,
  author       = {Pith},
  title        = {Pith review of: Retrieval Augmented Generation based Large Language Models for Causality Mining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3AJBO7H4}},
  note         = {Machine review of arXiv:2505.23944}
}
read the original abstract

Causality detection and mining are important tasks in information retrieval due to their enormous use in information extraction, and knowledge graph construction. To solve these tasks, in existing literature there exist several solutions -- both unsupervised and supervised. However, the unsupervised methods suffer from poor performance and they often require significant human intervention for causal rule selection, leading to poor generalization across different domains. On the other hand, supervised methods suffer from the lack of large training datasets. Recently, large language models (LLMs) with effective prompt engineering are found to be effective to overcome the issue of unavailability of large training dataset. Yet, in existing literature, there does not exist comprehensive works on causality detection and mining using LLM prompting. In this paper, we present several retrieval-augmented generation (RAG) based dynamic prompting schemes to enhance LLM performance in causality detection and extraction tasks. Extensive experiments over three datasets and five LLMs validate the superiority of our proposed RAG-based dynamic prompting over other static prompting schemes.

Figures

Figures reproduced from arXiv: 2505.23944 by the authors.

Figure 1
Figure 1. Logarithmic of the number of causal connec [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Dynamic fewshot selection mechanism of Pattern RAG and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Plot of the number of examples vs. perfor [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Causality Extraction Prompt A Dataset Discription SemEval: For our causality detection task, we se￾lect 249 causal sentences from the test set that are absent in our Fewshot Example DB and a random sample of 300 non-causal sentences from the test data (9). We label the…
Figure 5
Figure 5. Figure 5: Causality Detection Prompt [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Causal Connective Extraction Prompt Dataset Split #total unique instances #causal relations #non-causal relations Fewshot Ex￾ample DB - 2,365 2,365 - SemEval Train 8,000 1,003 6,997 Test 2,717 328 2,389 Task1 input 549 249 300 Task2 input 249 249 - Train 2,000 2,000 - …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 32 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

  4. [4]

    Wajid Ali, Wanli Zuo, Rahman Ali, Xianglin Zuo, and Gohar Rahman. 2021. Causality mining in natural languages using machine and deep learning techniques: A survey. Applied Sciences, 11(21):10064

  5. [5]

    Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. 2023. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 1

  6. [6]

    Steven Bethard and James H Martin. 2008. Learning semantic links from a corpus of parallel temporal and causal relations. In Proceedings of ACL-08: HLT, Short Papers, pages 177--180

  7. [7]

    Eduardo Blanco, Nuria Castell, and Dan I Moldovan. 2008. Causal relation extraction. In Lrec, volume 66, page 74

  8. [8]

    Tom B Brown. 2020. Language models are few-shot learners. arXiv preprint ArXiv:2005.14165

Show all 50 references
  1. [9]

    Chih-Yao Chen and Cheng-Te Li. 2021. Zs-bert: Towards zero-shot relation extraction with attribute representation learning. arXiv preprint arXiv:2104.04697

  2. [10]

    Tirthankar Dasgupta, Rupsa Saha, Lipika Dey, and Abir Naskar. 2018. Automatic extraction of causal relations from text using linguistically informed deep neural networks. In Proceedings of the 19th annual SIGdial meeting on discourse and dialogue, pages 306--316

  3. [11]

    Tharini N De Silva, Xiao Zhibo, Zhao Rui, and Mao Kezhi. 2017. Causal relation identification using convolutional neural networks and knowledge based features. International Journal of Computer and Systems Engineering, 11(6):696--701

  4. [12]

    Sefika Efeoglu and Adrian Paschke. 2024. Retrieval-augmented generation-based relation extraction. arXiv preprint arXiv:2404.13397

  5. [13]

    Daniela Garcia, EDF-DER, and IMA-TIEM. 1997. Coatis, an nlp system to locate expressions of actions connected by causality links. In International Conference on Knowledge Engineering and Knowledge Management, pages 347--352. Springer

  6. [14]

    Roxana Girju. 2003. Automatic detection of causal relations for question answering. In Proceedings of the ACL 2003 workshop on Multilingual summarization and question answering, pages 76--83

  7. [15]

    Roxana Girju, Dan I Moldovan, et al. 2002. Text mining for causal relations. In FLAIRS, volume 2, pages 360--364

  8. [16]

    Goddu and A Gopnik

    M.K. Goddu and A Gopnik. 2024. The development of human causal learning and reasoning. Nature Review Psychology, 3:319--339

  9. [17]

    Harsha Gurulingappa, Abdul Mateen Rajput, Angus Roberts, Juliane Fluck, Martin Hofmann-Apitius, and Luca Toldo. 2012. Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports. Journal of biomedical informat...

  10. [18]

    Chikara Hashimoto, Kentaro Torisawa, Julien Kloetzer, Motoki Sano, Istv \'a n Varga, Jong-Hoon Oh, and Yutaka Kidawara. 2014. Toward future scenario generation: Extracting event causality exploiting semantic relation, context, and association features. In Proceedings of the 52...

  11. [19]

    Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid O S \'e aghdha, Sebastian Pad \'o , Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2019. Semeval-2010 task 8: Multi-way classification of semantic relations between pairs of nominals. arXiv prepr...

  12. [20]

    Christopher Hidey and Kathleen McKeown. 2016. Identifying causal relations using parallel wikipedia articles. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1424--1433

  13. [21]

    Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional lstm-crf models for sequence tagging. arXiv preprint arXiv:1508.01991

  14. [22]

    Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. 2023. Mistral 7b. arXiv preprint arXiv:2310.06825

  15. [23]

    Zhijing Jin, Jiarui Liu, Zhiheng Lyu, Spencer Poff, Mrinmaya Sachan, Rada Mihalcea, Mona Diab, and Bernhard Sch \"o lkopf. 2023. Can large language models infer causation from correlation? arXiv preprint arXiv:2306.05836

  16. [24]

    Leo Joskowicz, T Ksiezyck, and Ralph Grishman. 1989. Deep domain models for discourse analysis. In 1989 The Annual AI Systems in Government Conference, pages 195--196. IEEE Computer Society

  17. [25]

    Md Ahsanul Kabir. 2023. Deep Learning Based Methods for Automatic Extraction of Syntactic Patterns and their Application for Knowledge Discovery. Ph.D. thesis

  18. [26]

    Randy M Kaplan and Genevieve Berry-Rogghe. 1991. Knowledge-based acquisition of causal relationships in text. Knowledge Acquisition, 3(3):317--337

  19. [27]

    Humayun Kayesh, Md Saiful Islam, and Junhu Wang. 2019. On event causality detection in tweets. arXiv preprint arXiv:1901.03526

  20. [28]

    Christopher SG Khoo, Jaklin Kornfilt, Robert N Oddy, and Sung Hyon Myaeng. 1998. Automatic extraction of cause-effect information from newspaper text without knowledge-based inferencing. Literary and linguistic computing, 13(4):177--186

  21. [29]

    John Kontos and Maria Sidiropoulou. 1991. On the acquisition of causal knowledge from scientific texts with attribute grammars. International Journal of Applied Expert Systems, 4(1):31--48

  22. [30]

    Yuquan Lan, Dongxu Li, Yunqi Zhang, Hui Zhao, and Gang Zhao. 2023. Modeling zero-shot relation classification as a multiple-choice problem. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1--8. IEEE

  23. [31]

    Xiaonan Li, Kai Lv, Hang Yan, Tianyang Lin, Wei Zhu, Yuan Ni, Guotong Xie, Xiaoling Wang, and Xipeng Qiu. 2023. Unified demonstration retriever for in-context learning. arXiv preprint arXiv:2305.04320

  24. [32]

    Zhaoning Li, Qi Li, Xiaotian Zou, and Jiangtao Ren. 2021. Causality extraction based on self-attentive bilstm-crf with transferred embeddings. Neurocomputing, 423:207--219

  25. [33]

    Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. 2021. What makes good in-context examples for gpt- 3 ? arXiv preprint arXiv:2101.06804

  26. [34]

    Junbao Liu, Xizhong Qin, Xiaoqin Ma, and Wensheng Ran. 2023. Freda: Few-shot relation extraction based on data augmentation. Applied Sciences, 13(14):8312

  27. [35]

    Harsha Nori, Yin Tat Lee, Sheng Zhang, Dean Carignan, Richard Edgar, Nicolo Fusi, Nicholas King, Jonathan Larson, Yuanzhi Li, Weishung Liu, et al. 2023. Can generalist foundation models outcompete special-purpose tuning? case study in medicine. arXiv preprint arXiv:2311.16452

  28. [36]

    Partha Pakray and Alexander Gelbukh. 2014. An open-domain cause-effect relation detection from paired nominals. In Mexican International Conference on Artificial Intelligence, pages 263--271. Springer

  29. [37]

    Santanu Pal, Partha Pakray, Dipankar Das, and Sivaji Bandyopadhyay. 2010. Ju: a supervised approach to identify semantic relations from paired nominals. In Proceedings of the 5th International Workshop on Semantic Evaluation, pages 206--209

  30. [38]

    Edoardo Maria Ponti and Anna Korhonen. 2017. Event-related features in feedforward neural networks contribute to identifying causal relations in discourse. In Proceedings of the 2nd Workshop on Linking Models of Lexical, Sentential and Discourse-level Semantics, pages 25--30

  31. [39]

    Kira Radinsky, Sagie Davidovich, and Shaul Markovitch. 2012. Learning causality for news events prediction. In Proceedings of the 21st international conference on World Wide Web, pages 909--918

  32. [40]

    Mehwish Riaz and Roxana Girju. 2010. Another look at causality: Discovering scenario-specific contingency relationships with no supervision. In 2010 IEEE Fourth International Conference on Semantic Computing, pages 361--368. IEEE

  33. [41]

    Jawad Sadek. 2013. Automatic detection of arabic causal relations. In Natural Language Processing and Information Systems: 18th International Conference on Applications of Natural Language to Information Systems, NLDB 2013, Salford, UK, June 19-21, 2013. Proceedings 18, pages ...

  34. [42]

    Craig Silverstein, Sergey Brin, Rajeev Motwani, and Jeff Ullman. 2000. Scalable techniques for mining causal structures. Data Mining and Knowledge Discovery, 4:163--192

  35. [43]

    Antonio Sorgente, Giuseppe Vettigli, and Francesco Mele. 2013. Automatic extraction of cause-effect relations in natural language text. DART@ AI* IA, 2013:37--48

  36. [44]

    Yuanhe Tian, Yan Song, and Fei Xia. 2022. Improving relation extraction through syntax-induced pre-training with dependency masking. In Findings of the Association for Computational Linguistics: ACL 2022, pages 1875--1886

  37. [45]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

  38. [46]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Learning to retrieve in-context examples for large language models. arXiv preprint arXiv:2307.07164

  39. [47]

    Shanchan Wu and Yifan He. 2019. Enriching pre-trained language model with entity information for relation classification. In Proceedings of the 28th ACM international conference on information and knowledge management, pages 2361--2364

  40. [48]

    Yiming Zhang, Shi Feng, and Chenhao Tan. 2022. Active example selection for in-context learning. arXiv preprint arXiv:2211.04486

  41. [49]

    Yuzhe Zhang, Yipeng Zhang, Yidong Gan, Lina Yao, and Chen Wang. 2024. Causal graph discovery with retrieval-augmented generation based large language models. arXiv preprint arXiv:2402.15301

  42. [50]

    Yi Zhao, Huaiyu Wan, Jianwei Gao, and Youfang Lin. 2019. Improving relation classification by entity pair graph. In Asian Conference on Machine Learning, pages 1156--1171. PMLR

Pith tools

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