Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Reasoning with RAGged events: RAG-Enhanced Event Knowledge Base Construction and reasoning with proof-assistants

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

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

Pith's one-line read The paper's central claim is an 'inverse calibration' effect: external knowledge enhancement helps weak AI models but degrades strong ones during event extraction from historical text.

desk verdict A promising but under-powered pilot: the inverse calibration principle is not established by one small table, though the Coq translation and the three-strategy comparison are worth a look. read the letter →

arxiv 2506.07042 v3 pith:23EHBQBO submitted 2025-06-08 cs.CL

Stergios Chatzikyriakidis This is my paper
classification cs.CL
keywords historicaleventextractionretrieval-augmentedgenerationlargelanguagemodelsRDFknowledgegraphsCoqproofassistantinversecalibrationprincipleontologydiscoveryThucydides
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 argues that adding external knowledge to large language models — through knowledge-graph lookups or retrieval-augmented generation (RAG) — is not a universal improvement for automatic historical event extraction. Using the first nine chapters of Thucydides, it compares GPT-4o, Claude-3.5, and Llama 3.2 under three conditions: pure base generation, knowledge-graph enhancement, and RAG. The empirical result is an 'inverse calibration principle': the strongest models extract the most events (36 and 39) with no augmentation, while the weaker Llama jumps from 10 to 37-38 events under simple RAG but collapses to 6 and 0 events under more complex RAG. The paper further claims that RAG-discovered event types absent from DBpedia — such as MigrationEvent and WarEvent — are legitimate domain structures, and that translating RDF output into Coq proof-assistant specifications allows higher-order reasoning such as multi-step causal proofs and BCE date arithmetic. The relevance is that RAG system design should be calibrated to model strength and task goals rather than assuming more retrieval is better.

What carries the argument

The paper's two load-bearing mechanisms are the 'inverse calibration principle' — the named generalization that enhancement benefit varies inversely with model capability, with stronger models best served by no augmentation and weaker models needing carefully bounded retrieval — and the automated RDF-to-Coq translation pipeline, which maps extracted RDF/Turtle events into Coq inductive types and dependent types, enabling temporal arithmetic on BCE dates and formal multi-step causal proofs. The pipeline's four stages (ontological discovery validation, type-system construction, relationship formalization, and theorem generation) are what convert RAG-discovered event classes into verifiable semantic structures.

What would settle it

A controlled replication with several runs per condition on a different well-documented historical corpus: if a strong model's event coverage increases under RAG, or if Llama 3.2's performance does not first rise then collapse as RAG complexity grows, the inverse calibration principle would be falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes two connected results. First, enhancement-strategy effectiveness inversely correlates with model capability: for knowledge-rich models (GPT-4o, Claude-3.5), pure base generation yields maximum event coverage with minimal hallucinations, while RAG and knowledge-graph injection reduce event counts and introduce errors such as misattributed battle coordinates and temporal conflation. For the smaller Llama 3.2, simple RAG configurations are strongly beneficial (10 to 37-38 events), but increasing RAG complexity pushes it over a fragility threshold into systematic collapse (6 and 0 events). Second, the apparent 'hallucinations' in RAG output — event types like ET_MigrationEvent and ET_WarEvent that violate DBpedia's taxonomy — are argued to be legitimate domain-specific semantic structures, validated by their smooth integration into Coq's type system and by formally proven theorems about historical causation. The paper presents the RDF-to-Coq translation pipeline as the mechanism that moves historical analysis from first-order graph reasoning to higher-order formal verification.

Load-bearing premise

The load-bearing premise is that a single evaluation run on the first nine chapters of Thucydides, with qualitative manual precision labels, is representative enough to support a general 'inverse calibration principle' about model capability and RAG effectiveness.

Editorial extensions

If this is right

  • In knowledge-rich domains, base generation should be the default for strong models; RAG should be reserved for cases requiring coordinate precision or metadata completeness, not event coverage.
  • For weaker models, simple RAG configurations can be highly beneficial, but retrieval complexity must be kept low; beyond a fragility threshold, performance collapses.
  • RDF/OWL knowledge graphs are insufficient for historical reasoning tasks such as multi-step causation and counterfactuals; translating to a proof assistant like Coq adds formal verification capabilities.
  • Evaluation of RAG event extraction should treat domain-specific event types as potential discoveries rather than automatic errors against general-purpose ontologies.
  • Strategy selection should be driven by the application's target dimension (coverage vs. precision) rather than by a blanket 'more knowledge is better' assumption.

Reading between the lines

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

  • A testable extension: running the same three-condition comparison on another well-documented historical corpus (e.g., Caesar's Gallic Wars) should reproduce the pattern — strong models' event counts flat or declining under RAG, and a small model showing an inverted-U — which would strengthen the principle beyond a single corpus.
  • If the inverse calibration holds generally, RAG system design should first estimate a model's internal knowledge density for the target domain (base-generation coverage is a cheap proxy) and set retrieval complexity accordingly, rather than defaulting to maximal retrieval.
  • The Coq-validation method suggests a reusable pattern for ontology engineering: treat extraction outliers as candidate domain classes and use a proof assistant to test whether they support consistent, inferentially useful structures before adopting them into an ontology.
  • The precision-over-coverage trade-off implies that for applications like digital humanities scholarship, the right strategy depends on the question: event catalogs want base generation, while prosopography-style metadata wants bounded RAG.
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 a two-phase pipeline for historical event extraction from Thucydides: Phase 1 compares base generation, knowledge-graph enhancement, and RAG across GPT-4o, Claude-3.5, and Llama 3.2; Phase 2 converts the extracted RDF representations into Coq specifications for higher-order temporal and causal reasoning. The central empirical claim is an 'inverse calibration principle': stronger models perform best with no external enhancement, while weaker models improve under simple RAG but collapse under complex RAG. The paper also claims that RAG-discovered event types such as WarEvent and MigrationEvent are legitimate domain-specific structures, validated by their representability in Coq. The conclusions generalize these findings to a general principle about model capability, enhancement intensity, and domain characteristics.

Significance. If the empirical claims were supported, the inverse calibration principle would be a useful counterpoint to the common assumption that more retrieval always helps, and the Coq translation would be an interesting bridge between LLM extraction and formal reasoning. The paper also makes a sensible observation about the granularity mismatch between general-purpose ontologies and domain-specific event types. However, the evidence is currently too thin: the central empirical principle rests on a single run per condition on a small corpus with qualitative precision labels, the RAG configurations are underspecified, and the Coq validation is largely circular. Credit is due for the reproducible data/code link, the use of deterministic decoding, and a focused related-work discussion, but these do not compensate for the load-bearing weaknesses.

major comments (4)
  1. [§3.2, Table 1] The inverse calibration principle is inferred from a single run for each model and configuration on the first 9–10 chapters of Thucydides, with event counts as single integers and precision labels assigned qualitatively by the author. The Llama trajectory (10 → 38 → 37 → 6 → 0) is presented as a systematic fragility threshold, but with no repeated trials, confidence intervals, or significance tests it cannot be distinguished from retrieval-set luck, API non-determinism, or a single malformed prompt. This is load-bearing because the paper's central claim is that enhancement effectiveness inversely correlates with model capability; the current evidence is a descriptive reading of one table.
  2. [§4.4, Stages 1 and 4] The Coq validation of RAG-discovered event types is circular. Stage 1 defines valid event types as those with domain-specific terms and contextual coherence (temporal consistency, specific participants, geographical anchoring), and the rag_discoveries_coherent theorem then proves that RAG-extracted events with 'proper temporal, participant, and class structure' are coherent. Since the theorem's assumptions are exactly the criteria used to declare the events valid, the proof establishes consistency of the definition rather than showing that RAG discoveries are legitimate domain knowledge. Moreover, no actual Coq development, complete theorem statements, or proof scripts are included; the paper shows only illustrative definitions.
  3. [§3.1–§3.2] The RAG configurations labeled RAG 1–4 are never specified, so the central comparison cannot be reproduced or interpreted. The text gives chunk size and embedding model but omits k values, retrieved context sizes, and what distinguishes the four configurations; without this, the reported collapse of Llama under RAG 3–4 could equally be an implementation artifact. This also prevents any quantitative evaluation of the claim that coordinate precision and RDF completeness improve under RAG.
  4. [§3.2] The claim that RAG 'improves precision' is not supported by the data as reported. Precision is a qualitative label with no rubric, no inter-annotator agreement, and no numerical scores; the only quantitative metric, coordinate distance, is mentioned in §3.1 but no per-condition values are given. With event counts also unreplicated, the paper cannot substantiate the trade-off between coverage and precision that underlies the inverse calibration principle.
minor comments (5)
  1. [Abstract and §3] The abstract says '10 first chapters' of Thucydides while Section 3 says 'first 9 chapters'; the corpus size should be stated consistently.
  2. [Table 1] The table contains malformed model names such as 'Claude 4Base Generation' and inconsistent capitalization of 'LLAMA'; please clean the formatting and use uniform row separators.
  3. [Introduction and §5] The 'inverse calibration principle' is stated as an established finding in the Introduction and Conclusions, but it is not formulated as a falsifiable hypothesis or tested out-of-sample; consider presenting it as a tentative observation pending replication.
  4. [References] Reference [17] lacks publication details, and reference [35] contains a broken author string ('In the real Dai'); these should be corrected.
  5. [Figure 2] The RAG pipeline figure has only a caption-like label and no detailed parameter annotations; please add a proper caption and specify the components involved.

Circularity Check

1 steps flagged · score 6.0 of 10

Coq validation of RAG-discovered event types is definitional; the empirical RAG comparison is independent but based on single runs.

  1. self definitional [Section 4.4, Stage 4: Theorem Generation (also Stage 1: Ontological Discovery Validation)]
    "The rag_discoveries_coherent theorem formally proves that RAG-extracted events with proper temporal, participant, and class structure represent valid domain knowledge"

    Stage 1 defines what counts as a legitimate non-standard class using exactly these criteria: 'semantic indicators (presence of domain-specific terms like War, Migration, Siege) and contextual coherence (temporal consistency, specific participants, geographical anchoring)'. The Stage 4 theorem then 'proves' that events with proper temporal, participant, and class structure are coherent. This is true by construction: the theorem's preconditions are the paper's own definition of coherence. Therefore the claimed independent validation that RAG-discovered event types are legitimate semantic structures rather than hallucinations reduces to a definitional tautology, not to a formal discovery about the domain.

full rationale

The empirical comparison in Section 3.2 is not circular: the event counts, precision labels, and RDF issue notes in Table 1 are observations of LLM outputs under different strategies, and the 'inverse calibration principle' is a post-hoc summary of those observations rather than a prediction fitted to a subset and then validated on a closely related quantity. The paper's self-citations to prior Coq work on natural-language semantics are supporting context, not load-bearing reductions. The central circularity is in the formal-validation contribution: Section 4.4 first defines ontological legitimacy via semantic indicators and contextual coherence, then presents the rag_discoveries_coherent theorem as independent proof that RAG-discovered event types are legitimate. Since the theorem's assumptions are the definition of coherence, the proof cannot independently validate the ontological-discovery thesis. This makes the formal-validation half of the paper's contribution partially circular, while the empirical comparison retains independent content. The score reflects that partial circularity, not the empirical weaknesses (single runs, qualitative precision labels), which are correctness risks rather than circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The empirical 'inverse calibration' claim rests on the choice of a single corpus and manual evaluation criteria; the Coq validation rests on author-defined coherence criteria and the assumption that representability in Coq implies semantic legitimacy.

free parameters (2)
  • RAG configuration complexity = Four unnamed configurations (RAG1-4), exact k and source-combination values not reported
    The paper reports the choice of which RAG level succeeds or fails for Llama, but does not disclose the exact k values or retrieval composition, making the central performance curve unfalsifiable.
  • Text chunk size = 1500-2000 tokens with 200-character overlap
    Chosen manually; affects retrieval and extraction, but no sensitivity analysis is provided.
assumptions (4)
  • standard math Coq's inductive type theory and the Z library are appropriate for encoding temporal points and event relations.
    Used in Section 4.3-4.4 for TimePoint, is_bc_event, and cascading_causation.
  • domain assumption DBpedia, Wikidata, ConceptNet, and the LACRIMALit ontology provide reliable external knowledge for historical entity and location resolution.
    Invoked in Section 3 as the knowledge sources for enhancement and RAG.
  • ad hoc to paper Semantic indicators (presence of domain terms like 'War' or 'Migration') and contextual coherence (temporal consistency, specific participants, geographical anchoring) can distinguish legitimate event types from hallucinations.
    Defined by the authors in Section 4.4 Stage 1 and applied to deem RAG-discovered types valid.
  • ad hoc to paper Representability in Coq's type system is sufficient to establish semantic legitimacy of an event type.
    Section 4.4 states that integration of RAG-discovered types into Coq 'validates' their legitimacy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reasoning with RAGged events: RAG-Enhanced Event Knowledge Base Construction and reasoning with proof-assistants." pith.science (2026). https://pith.science/paper/23EHBQBO

@misc{pith2026250607042,
  author       = {Pith},
  title        = {Pith review of: Reasoning with RAGged events: RAG-Enhanced Event Knowledge Base Construction and reasoning with proof-assistants},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/23EHBQBO}},
  note         = {Machine review of arXiv:2506.07042}
}
read the original abstract

Extracting structured computational representations of historical events from narrative text remains computationally expensive when constructed manually. While RDF/OWL reasoners enable graph-based reasoning, they are limited to fragments of first-order logic, preventing deeper temporal and semantic analysis. This paper addresses both challenges by developing automatic historical event extraction models using multiple LLMs (GPT-4, Claude, Llama 3.2) with three enhancement strategies: pure base generation, knowledge graph enhancement, and Retrieval-Augmented Generation (RAG). We conducted comprehensive evaluations using historical texts from Thucydides. Our findings reveal that enhancement strategies optimize different performance dimensions rather than providing universal improvements. For coverage and historical breadth, base generation achieves optimal performance with Claude and GPT-4 extracting comprehensive events. However, for precision, RAG enhancement improves coordinate accuracy and metadata completeness. Model architecture fundamentally determines enhancement sensitivity: larger models demonstrate robust baseline performance with incremental RAG improvements, while Llama 3.2 shows extreme variance from competitive performance to complete failure. We then developed an automated translation pipeline converting extracted RDF representations into Coq proof assistant specifications, enabling higher-order reasoning beyond RDF capabilities including multi-step causal verification, temporal arithmetic with BC dates, and formal proofs about historical causation. The Coq formalization validates that RAG-discovered event types represent legitimate domain-specific semantic structures rather than ontological violations.

Figures

Figures reproduced from arXiv: 2506.07042 by the authors.

Figure 1
Figure 1. LLM System Prompt for Historical Event Extraction Knowledge Enhanced (No-RAG): Here, we use knowledge integration but no RAG. We at￾tempt entity extraction, knowledge graph retrieval queries to Wikidata, DBpedia, and ConceptNet [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. The RAG Pipeline 3.2 Enhancement Strategy Comparison Our systematic investigation examined how different knowledge integration approaches affect ex￾traction performance across model architectures. We evaluated the three enhancement strategies using identical prompting frameworks to ensure fair comparison. The results are shown in [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. RDF/OWL capabilities versus historical reasoning requirements. The reasoning gap between limited RDF expressivity and complex historical analysis necessitates translation to formal verification systems. It is true that extensions such as SWRL or DL-safe Horn rules can cover additional entailed facts beyond pure RDFS. However, these rule regimes remain essentially first-order, lacking facilities for dependent types, … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 31 canonical work pages

  1. [1]

    Retrieval-based language models and appli- cations

    Akari Asai, Sewon Min, Zexuan Zhong, and Danqi Chen. Retrieval-based language models and appli- cations. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics: Tutorial Abstracts, pages 1–50, 2023

    work page 2023

  2. [2]

    Knowledge-Augmented Language Model Prompting for Zero-Shot Knowledge Graph Question Answering

    Jinheon Baek, Alham Fikri Aji, and Amir Saffari. Knowledge-Augmented Language Model Prompting for Zero-Shot Knowledge Graph Question Answering. InProceedings of the First Workshop on Matching From Unstructured and Structured Data (MATCHING 2023), pages 70–98, Toronto, ON, Canada, 2023. Association for Computational Linguistics

    work page 2023

  3. [3]

    R. C. Barron, V. Grantcharov, S. Wanna, M. E. Eren, M. Bhattarai, N. Solovyev, G. Tompkins, C. Nicholas, K. Ø. Rasmussen, C. Matuszek, et al. Domain-Specific Retrieval-Augmented Generation Using Vector Stores, Knowledge Graphs, and Tensor Factorization. In2024 International Conference on Machine Learning and Applications (ICMLA), pages 1669–1676, 2024

    work page 2024

  4. [4]

    In: Proceedings of the 14th International Conference on Computational Semantics (IWCS)

    Bernardy, J.-P., Chatzikyriakidis, S.: Applied temporal analysis: A complete run of the FraCaS test suite. In: Proceedings of the 14th International Conference on Computational Semantics (IWCS). pp. 11–20 (2021)

    work page 2021

  5. [5]

    Rae, Erich Elsen, and Laurent Sifre

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George van den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, Diego de Las Casas, Aurelia Guy, Jacob Menick, Roman Ring, Tom Hennigan, Saffron Huang, Loren Maggiore, Chris Jones, Albin Cassirer, Andy Brock, Michela Paganini, Geoffrey Irving, Ori...

    work page 2022

  6. [6]

    Applied temporal analysis: A complete run of the FraCaS test suite

    Jean-Philippe Bernardy and Stergios Chatzikyriakidis. Applied temporal analysis: A complete run of the FraCaS test suite. InProceedings of the 14th International Conference on Computational Semantics (IWCS), pages 11–20, 2021

    work page 2021

  7. [7]

    Og-rag: Ontology-grounded retrieval-augmented generation for large language models.arXiv preprint arXiv:2412.15235, 2024

    Kartik Sharma, Peeyush Kumar, and Yunqing Li. Og-rag: Ontology-grounded retrieval-augmented generation for large language models.arXiv preprint arXiv:2412.15235, 2024

    arXiv 2024

  8. [8]

    Proof assistants for natural language semantics

    Stergios Chatzikyriakidis and Zhaohui Luo. Proof assistants for natural language semantics. InLogical Aspects of Computational Linguistics. Celebrating 20 Years of LACL (1996–2016) 9th International Conference, LACL 2016, Nancy, France, December 5-7, 2016, Proceedings 9, pages 85–98. Springer, 2016

    work page 1996

Show all 37 references
  1. [9]

    The power of noise: Redefining retrieval for RAG systems

    Florin Cuconasu, Giovanni Trappolini, Federico Siciliano, Simone Filice, Cesare Campagnano, Yoelle Maarek, Nicola Tonellotto, and Fabrizio Silvestri. The power of noise: Redefining retrieval for RAG systems. InProceedings of the 47th International ACM SIGIR Conference on Resea...

    2024

  2. [10]

    Event extraction by answering (almost) natural questions

    Xinya Du and Claire Cardie. Event extraction by answering (almost) natural questions. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 671– 683, 2020

    2020

  3. [11]

    Retrieval-augmented generative question answering for event argument extrac- tion

    Xinya Du and Heng Ji. Retrieval-augmented generative question answering for event argument extrac- tion. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4649–4666, 2022

    2022

  4. [12]

    Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024. Title Suppressed Due to Excessive Length 15

    2024

  5. [13]

    Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2023

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen Wang. Retrieval-augmented generation for large language models: A survey.arXiv preprint arXiv:2312.10997, 2023

    2023 arXiv

  6. [14]

    Demonstration retrieval- augmented generative event argument extraction

    Shiming He, Yu Hong, Shuai Yang, Jianmin Yao, and Guodong Zhou. Demonstration retrieval- augmented generative event argument extraction. InProceedings of the 2024 Joint International Con- ference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024...

    2024

  7. [15]

    Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequeda, Steffen Staab, and Antoine Zimmermann

    Aidan Hogan, Eva Blomqvist, Michael Cochez, Claudia d’Amato, Gerard de Melo, Claudio Gutierrez, Sabrina Kirrane, José Emilio Labra Gayo, Roberto Navigli, Sebastian Neumaier, Axel-Cyrille Ngonga Ngomo, Axel Polleres, Sabbir M. Rashid, Anisa Rula, Lukas Schmelzeisen, Juan Sequed...

    2021

  8. [16]

    Leveraging passage retrieval with generative models for open domain question answering

    Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 874–880, 2021

    2021

  9. [17]

    Long-ContextLLMsMeetRAG:Overcoming Challenges for Long Inputs in RAG.arXiv preprint arXiv:2410, 2024

    BowenJin,JinsungYoon,JiaweiHan,andSercanO.Arik. Long-ContextLLMsMeetRAG:Overcoming Challenges for Long Inputs in RAG.arXiv preprint arXiv:2410, 2024

    2024

  10. [18]

    Dense passage retrieval for open-domain question answering

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages...

    2020

  11. [19]

    Lehmann, R

    J. Lehmann, R. Isele, M. Jakob, A. Jentzsch, D. Kontokostas, P. N. Mendes, S. Hellmann, M. Morsey, P. Van Kleef, S. Auer, et al. DBpedia–a large-scale, multilingual knowledge base extracted from wikipedia. Semantic Web, 6(2):167–195, 2015

    2015

  12. [20]

    Retrieval-augmented generation for knowledge-intensive nlp tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks. InAdvances in Neural ...

    2020

  13. [21]

    Intra-Event and Inter-Event Dependency-Aware Graph Network for Event Argument Extraction

    Hao Li, Yanan Cao, Yubing Ren, Fang Fang, Lanxue Zhang, Yingjie Li, and Shi Wang. Intra-Event and Inter-Event Dependency-Aware Graph Network for Event Argument Extraction. InFindings of the As- sociation for Computational Linguistics: EMNLP 2023, pages 6362–6372, Singapore, 20...

    2023

  14. [22]

    Jasper Linders and Jakub M. Tomczak. Knowledge graph-extended retrieval augmented generation for question answering. arXiv preprint arXiv:2504.08893, 2025

    2025 arXiv

  15. [23]

    RAG-based temporal reasoning for event relation extraction

    Jiawei Liu, Chunyu Kit, and Gina-Anne Levow. RAG-based temporal reasoning for event relation extraction. InProceedings of the AAAI Conference on Artificial Intelligence, pages 13245–13253, 2023

    2023

  16. [24]

    HyperGraphRAG: Retrieval-Augmented Generation with Hypergraph-Structured Knowledge Representation.arXiv preprint arXiv:2503, 2025

    Haoran Luo, Guanting Chen, Yandan Zheng, Xiaobao Wu, Yikai Guo, Qika Lin, Yu Feng, Zemin Kuang, Meina Song, Yifan Zhu, and others. HyperGraphRAG: Retrieval-Augmented Generation with Hypergraph-Structured Knowledge Representation.arXiv preprint arXiv:2503, 2025

    2025

  17. [25]

    Few-shot event detection: An empirical study and a unified view

    Yubo Ma, Zehao Wang, Yixin Cao, and Aixin Sun. Few-shot event detection: An empirical study and a unified view. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 2845–2857, 2023

    2023

  18. [26]

    Fine-tuning or retrieval? comparing knowledge injection in llms.arXiv preprint arXiv:2312.05934, 2023

    Oded Ovadia, Menachem Brief, Moshik Mishaeli, and Oren Elisha. Fine-tuning or retrieval? comparing knowledge injection in llms.arXiv preprint arXiv:2312.05934, 2023

    2023 arXiv

  19. [27]

    The LACRIMALit Ontology of Crisis: An Event-Centric Model for Digital History.Information, 13(8):398, 2022

    Maria Papadopoulou, Christophe Roche, and Eleni-Melina Tamiolaki. The LACRIMALit Ontology of Crisis: An Event-Centric Model for Digital History.Information, 13(8):398, 2022

    2022

  20. [28]

    Knowledge graph refinement: A survey of approaches and evaluation methods.Se- mantic Web, 8(3):489–508, 2017

    Heiko Paulheim. Knowledge graph refinement: A survey of approaches and evaluation methods.Se- mantic Web, 8(3):489–508, 2017

    2017

  21. [29]

    W. Su, Y. Tang, Q. Ai, Z. Wu, and Y. Liu. DRAGIN: Dynamic Retrieval Augmented Generation based on the Information Needs of Large Language Models.arXiv preprint arXiv:2403.10081, 2024

    2024 arXiv

  22. [30]

    RA- NER: Retrieval augmented NER for knowledge intensive named entity recognition.Amazon Science Publications, 2024

    Xinyu Wang, Yong Jiang, Nguyen Bach, Tao Wang, Zhongqiang Huang, Fei Huang, and Kewei Tu. RA- NER: Retrieval augmented NER for knowledge intensive named entity recognition.Amazon Science Publications, 2024. 16 S. Chatzikyriakidis

    2024

  23. [31]

    Retrieval-augmented event argument extraction

    Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Retrieval-augmented event argument extraction. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics, pages 2156–2168, 2023

    2023

  24. [32]

    How faithful are RAG models? Quantifying the tug-of-war between RAG and LLMs’ internal prior.arXiv preprint arXiv:2404, 2024

    Kevin Wu, Eric Wu, and James Zou. How faithful are RAG models? Quantifying the tug-of-war between RAG and LLMs’ internal prior.arXiv preprint arXiv:2404, 2024

    2024

  25. [33]

    Manning, Percy Liang, and Jure Leskovec

    Michihiro Yasunaga, Antoine Bosselut, Hongyu Ren, Xikun Zhang, Christopher D. Manning, Percy Liang, and Jure Leskovec. Deep bidirectional language-knowledge graph pretraining. InAdvances in Neural Information Processing Systems, volume 35, pages 14951–14964, 2022

    2022

  26. [34]

    Retrieval-augmented generation across heterogeneous knowledge

    Wenhao Yu, Dan Iter, Shuohang Wang, Yichong Xu, Mingxuan Ju, Soumya Sanyal, Chenguang Zhu, Michael Zeng, and Meng Jiang. Retrieval-augmented generation across heterogeneous knowledge. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Co...

    2022

  27. [35]

    RA-NER: Retrieval augmented NER for knowledge intensive named entity recognition

    In the real Dai, Chen Luo, Zhen Li, Xianfeng Tang, Hanqing Lu, Rahul Goutam, and Haiyang Zhang. RA-NER: Retrieval augmented NER for knowledge intensive named entity recognition. InThe Second Tiny Papers Track at ICLR 2024, 2024

    2024

  28. [36]

    Retrieval-augmented generation for ai-generated content: A survey

    Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang, Wentao Zhang, Jie Jiang, and Bin Cui. Retrieval-augmented generation for ai-generated content: A survey. arXiv preprint arXiv:2402.19473, 2024

    2024 arXiv

  29. [37]

    Event temporal relation extraction based on retrieval-augmented on llms.arXiv preprint arXiv:2403.15273, 2024

    Bingfeng Zhou, Yangqiu Song, and Kam-Fai Wong. Event temporal relation extraction based on retrieval-augmented on llms.arXiv preprint arXiv:2403.15273, 2024

    2024 arXiv

Pith tools

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