Pith. sign in

REVIEW 3 major objections 4 minor 47 references

Position Paper: Metadata Enrichment Model: Integrating Neural Networks and Semantic Knowledge Graphs for Cultural Heritage Applications

T0 review · 3 major / 4 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 This paper argues that an iterative detect-describe-decide loop can uncover nested features in cultural heritage images and enrich metadata accordingly.

desk verdict The dataset is real, but the paper's key novelty—the MVM recursion—is never actually run, so the central claim rests on a pipeline that exists only on paper. read the letter →

arxiv 2505.23543 v1 pith:XP4QUZM3 submitted 2025-05-29 cs.CV

Jan Ignatowicz , Krzysztof Kutt , Grzegorz J. Nalepa This is my paper
classification cs.CV
keywords metadataenrichmentculturalheritageknowledgegraphlargelanguagemodelsiterativevisualanalysisobjectdetectionincunabulalinkeddata
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes the Metadata Enrichment Model (MEM), a framework that combines fine-tuned computer vision models, large language models, and semantic knowledge graphs to enrich metadata for digitized cultural heritage objects. Its central innovation is the Multilayer Vision Mechanism (MVM), an iterative process that zooms into nested features, such as text inside seals or images inside stamps, rather than treating each image as a flat collection of objects. The authors argue that this recursion produces richer, more interoperable metadata, and they support the claim with a proof-of-concept on incunabula from the Jagiellonian Digital Library, releasing a manually annotated dataset of 105 manuscript pages. If MEM works as described, it would let libraries and archives automatically document fine-grained historical details that current schemas miss, and link them to external knowledge bases. The paper is a position paper, so the main contribution is the framework and the dataset, with the MVM loop as the key mechanism to be tested further.

What carries the argument

The Multilayer Vision Mechanism (MVM) is the core component: an iterative detect-crop-describe-decide loop. Detection or segmentation models identify candidate regions; a vision-language model like BLIP-2 turns each cropped region into a text description; an instruction-following model like Mistral-7B answers a predefined query about whether the region contains text or other features requiring transcription; rule-based guards on pixel size, confidence, and iteration depth constrain the loop. By separating detection from decision logic, MVM allows the detection models to be swapped without changing the workflow, which is what makes the framework adaptable to different cultural heritage collections.

What would settle it

Run the MVM decision step on a random sample of cropped regions that human annotators have confirmed to contain transcribable text, and count how many receive a 'NO' answer; a materially high false-NO rate would mean the recursion systematically halts before discovering the nested text that the method is designed to find.

Watch

Extended reading notes

Core claim

The central claim is that the Multilayer Vision Mechanism (MVM) improves visual analysis by iteratively detecting nested features within cultural artifacts. In the proposed workflow, a detection model such as YOLOv11 or Detectron2 finds high-level elements like seals or paragraphs; each region is cropped and passed to a language model that generates a text description; another language model then makes a YES or NO decision, combined with rule-based checks, about whether deeper analysis is warranted. The loop repeats until no new semantically meaningful features are found or a maximum depth is reached. The paper reports that a preliminary study on digitized incunabula validated the usefulness of MEM, with both YOLOv11 and Detectron2 able to detect the annotated classes, and it releases the annotated dataset to support further research.

Load-bearing premise

The load-bearing premise is that the language model's image descriptions and YES/NO decisions are accurate enough to steer the recursion; the paper reports no accuracy for this decision step, so a wrong 'NO' could stop analysis of a nested feature before it is found.

Editorial extensions

If this is right

  • If MVM proves reliable, metadata records would capture not just the presence of a seal but also the text inside it, making such details searchable and queryable for the first time.
  • Enriched metadata structured as RDF and published via SPARQL endpoints would allow cross-institutional queries, so a seal in one library's manuscript could be linked to a Wikidata entry about its historical owner.
  • The released 105-page annotated dataset, with ten object categories, gives other groups a starting point for fine-tuning detection models on historical documents without building annotations from scratch.
  • MEM's modular design means an institution could adopt only the object-detection stage first and later add the LLM decision layer and knowledge graph, easing incremental implementation in GLAM settings.
  • The dynamic ontology expansion could automatically define new classes, such as mem:SealWithText, when MVM detects a previously unclassified feature, something static schemas cannot do.

Reading between the lines

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

  • The same recursive detect-describe-decide loop could generalize beyond manuscripts to other nested-document domains, such as maps with inset cartouches, heraldic shields with mottos, or scientific plates with embedded annotations.
  • The paper does not report accuracy for the LLM decision step; a natural next experiment is to benchmark that YES/NO gate against human-verified nested text, since a wrong 'NO' stops the recursion early and would silently drop metadata.
  • At 105 pages, the dataset is small for fine-tuning large detectors; combining it with transfer learning from generic document datasets might be necessary before the method reaches production quality in underfunded institutions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes the Metadata Enrichment Model (MEM), a conceptual framework for enriching cultural heritage metadata by combining retrained computer vision models, large language models, and semantic knowledge graphs. Its stated key innovation is the Multilayer Vision Mechanism (MVM), an iterative loop that detects nested features (e.g., text within seals) by cropping regions of interest, describing them with an image-to-text model such as BLIP-2, and using an LLM such as Mistral-7B to decide whether to continue iterating. The authors report a proof-of-concept study on 105 manuscript pages from the Jagiellonian Digital Library, train YOLOv11 and Detectron2 to recognize ten flat object categories, release the annotated dataset, and discuss challenges including domain-specific fine-tuning, ontology alignment, computational cost, and ethical considerations. The paper concludes that the preliminary results validated MEM, but the reported experiments do not include any MVM recursion.

Significance. If the MVM mechanism worked as described, it would address a real gap in cultural heritage metadata enrichment, where structural and nested visual features are rarely captured. The released dataset of 105 manually annotated incunabula pages is a tangible contribution that can support future work, and the paper usefully catalogs practical issues such as GPU costs, ontology alignment, and human-in-the-loop verification. However, the paper's central claim about MVM is not empirically supported: the case study implements only single-pass object detection, with no end-to-end MVM run, no measurement of the LLM decision step, and no quantitative detection metrics. The paper is best read as a position or proposal paper, and its usefulness depends on the community's interest in the proposed architecture and dataset rather than on demonstrated performance.

major comments (3)
  1. [III.C and IV.A/IV.C] The central claim that MVM improves visual analysis by dynamically detecting nested features is never instantiated in the reported case study. Section IV.A describes training YOLOv11 and Detectron2 on ten flat categories (Paragraph, Seal, Stamp, etc.), but no cropped region is passed to BLIP-2, no Mistral-7B decision is recorded, no second-level detection or OCR is demonstrated, and no nested-feature annotations are released. The statement in Section IV.C that 'preliminary results validated the usefulness of MEM' therefore refers at most to single-pass object detection, not to the mechanism that constitutes MEM's novelty. The authors should either provide an end-to-end MVM trace on at least a few pages, or explicitly reframe the paper as a proposal and remove the validation claim.
  2. [IV.A] The empirical evidence for the object-detection component is not quantified. The sentence 'Both models showed the ability to detect desired visual elements' is the only result reported, with no precision, recall, mAP, error bars, or comparison against a baseline. Because this is the only empirical support offered in the paper, the authors should at minimum report standard detection metrics on their held-out set, or clearly label the study as a qualitative demonstration rather than a validation.
  3. [III.C and V.D] The recursion's reliability rests on an unmeasured component: the LLM YES/NO decision that determines whether to zoom into a region. The paper lists hallucination mitigations (rule-based guards, fine-tuning on domain Q&A pairs) but reports no accuracy for this decision step. A wrong 'NO' stops analysis prematurely and a wrong 'YES' wastes computation or introduces noise, so the utility of MVM is not established without at least a small pilot evaluation of the decision step. The authors should either supply such a pilot or identify this as a key open research question in the limitations section.
minor comments (4)
  1. [Abstract and Introduction] There are several grammatical and typographical errors, including 'we argue, that' (Abstract), 'Base challanges' (Section IV.B), 'struturing' (Section IV.B), 'makins' (Section III.D), 'expantion' (Section III.F), and 'Prossible' (Section V.D). A thorough language edit is needed.
  2. [II.C] The reference to 'oLLama' in the LLM list appears to be a typo for 'Ollama'; please correct the capitalization in the text and the bibliography entry.
  3. [II.E] The description of reference [41] as 'improve structural segmentation in historical documents' does not match the cited work, which is titled 'Handwriting recognition in historical documents with multimodal llm.' Either the citation or the description needs correction.
  4. [IV.A] Figure 3's caption says the image shows 'the detected classes: initial, header, and paragraph,' but the surrounding text refers to 'objects detected by the retrained YOLOv11 model' and the figure is not referenced in the dataset description. Please clarify the figure caption and ensure it matches the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper proposes a conceptual framework and reports a proof-of-concept; no prediction reduces to a fitted parameter or to a load-bearing self-citation.

full rationale

The paper contains no fitted-parameter-to-prediction chain. Its core contribution, the Multilayer Vision Mechanism (Section III.C), is presented as a conceptual workflow, and the case study (Section IV.A) trains YOLOv11 and Detectron2 on ten flat object categories; no MVM recursion, LLM decision step, or nested-feature output is measured. The statement in Section IV.C that "preliminary results validated the usefulness of MEM" is therefore unsupported, but that is an evidence gap, not circularity: the validation claim is not obtained by construction from the training data, by definition, or by a self-citation. The only self-referential mention, the CHExRISH project in the acknowledgment and Section III.B, is background context and is not load-bearing. No uniqueness theorem, ansatz smuggled in via citation, or renamed known result is invoked to force the framework. The paper's central claim may be under-supported, especially because the LLM-based decision quality is unmeasured, but it does not reduce to its own inputs by definition or by fitted values.

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

The framework has no fitted constants, but it relies on several unverified domain assumptions: transferability of detectors trained on 105 pages, reliability of LLM-gated recursion, and semantic correctness of automatic ontology extension. No new physical entities are introduced; MEM and MVM are compositions of existing technologies.

free parameters (3)
  • maximum iteration depth = 3 layers (example)
    Chosen by hand in Section III.C as a stopping bound for MVM; controls how many nested elements are exposed and is not derived from data.
  • minimum pixel size threshold = not specified
    Listed as a rule-based stop criterion for MVM in Section III.C; would affect which small elements are kept, but no value is given and no sensitivity analysis is supplied.
  • confidence threshold = not specified
    MVM combines LLM decisions with confidence thresholds in Section III.C; the paper does not report thresholds or their effect on precision and recall.
assumptions (4)
  • domain assumption Fine-tuning YOLOv11 and Detectron2 on 105 pages yields models good enough to support MEM's metadata enrichment.
    Section IV.A states both models 'showed the ability' to detect elements, but no accuracy is reported, so the upstream quality that everything downstream depends on is assumed rather than demonstrated.
  • domain assumption LLM-generated descriptions and YES/NO decisions are reliable enough to guide the iterative vision loop.
    Section III.C makes the continuation decision depend on BLIP-2 descriptions and Mistral-7B answers; the paper only mentions hallucination mitigation as a proposed safeguard, not as a measured property.
  • domain assumption The ten manually defined categories are the semantically meaningful structure for incunabula metadata enrichment.
    Section IV.A introduces classes such as Paragraph, Ornament, Stain, Stamp; their completeness and inter-annotator agreement are not assessed.
  • domain assumption Automatically extending an RDF ontology from detected visual elements preserves semantic correctness.
    Section III.B claims MEM dynamically defines new classes like mem:DecorativeSeal; the correctness of such automatic extension is stated without evaluation or expert validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Position Paper: Metadata Enrichment Model: Integrating Neural Networks and Semantic Knowledge Graphs for Cultural Heritage Applications." pith.science (2026). https://pith.science/paper/XP4QUZM3

@misc{pith2026250523543,
  author       = {Pith},
  title        = {Pith review of: Position Paper: Metadata Enrichment Model: Integrating Neural Networks and Semantic Knowledge Graphs for Cultural Heritage Applications},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XP4QUZM3}},
  note         = {Machine review of arXiv:2505.23543}
}
read the original abstract

The digitization of cultural heritage collections has opened new directions for research, yet the lack of enriched metadata poses a substantial challenge to accessibility, interoperability, and cross-institutional collaboration. In several past years neural networks models such as YOLOv11 and Detectron2 have revolutionized visual data analysis, but their application to domain-specific cultural artifacts - such as manuscripts and incunabula - remains limited by the absence of methodologies that address structural feature extraction and semantic interoperability. In this position paper, we argue, that the integration of neural networks with semantic technologies represents a paradigm shift in cultural heritage digitization processes. We present the Metadata Enrichment Model (MEM), a conceptual framework designed to enrich metadata for digitized collections by combining fine-tuned computer vision models, large language models (LLMs) and structured knowledge graphs. The Multilayer Vision Mechanism (MVM) appears as the key innovation of MEM. This iterative process improves visual analysis by dynamically detecting nested features, such as text within seals or images within stamps. To expose MEM's potential, we apply it to a dataset of digitized incunabula from the Jagiellonian Digital Library and release a manually annotated dataset of 105 manuscript pages. We examine the practical challenges of MEM's usage in real-world GLAM institutions, including the need for domain-specific fine-tuning, the adjustment of enriched metadata with Linked Data standards and computational costs. We present MEM as a flexible and extensible methodology. This paper contributes to the discussion on how artificial intelligence and semantic web technologies can advance cultural heritage research, and also use these technologies in practice.

Figures

Figures reproduced from arXiv: 2505.23543 by the authors.

Figure 1
Figure 1. Metadata Enrichment Model (MEM) main flow. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. An example image from the created dataset. The image shows the [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Possible entry modalities for MEM. • Image-to-Text Conversion: A language model like BLIP￾2 generates a text description of the cropped region (e.g. “A circular seal with Latin text along the border”). • LLM-Conducted Decision: A language model like Mistral-7B evaluates the description with predefined queries, such as “Does this region contain text requir￾ing transcription?” The LLM’s output (YES or NO) is combined … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: An example image from the created dataset. The image shows the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Example created ontology structure based on found new metadata and provided metadata by institution. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 31 canonical work pages

  1. [1]

    van Hooland and R

    S. van Hooland and R. Verborgh,Linked Data for Libraries, Archives and Museums. Facet Publishing, Jun. 2014. [Online]. Available: http://amzn.to/UUfjfN

    work page 2014

  2. [2]

    Knowledge graphs for cultural heritage and digital humanities,

    V . de Boer, “Knowledge graphs for cultural heritage and digital humanities,” inProceedings of the 5th Workshop on AnalySis, Understanding and ProMotion of HeritAge Contents, ser. SUMAC ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. 3. [Online]. Available: https://doi.org/10.1145/3607542.3617354

    arXiv 2023

  3. [3]

    Dublin Core Metadata for Resource Discovery,

    S. Weibel, J. Kunze, C. Lagoze, and M. Wolf, “Dublin Core Metadata for Resource Discovery,” Internet Engineering Task Force (IETF), Tech. Rep. RFC 2413, 1998

    work page 1998

  4. [4]

    Marc 21 format for bibliographic data,

    R. Guenther, “Marc 21 format for bibliographic data,”Library of Congress, 2003. [Online]. Available: https://www.loc.gov/marc/ bibliographic/

    work page 2003

  5. [5]

    The CIDOC conceptual reference model: An ontological approach to semantic interoperability of metadata,

    M. Doerr, “The CIDOC conceptual reference model: An ontological approach to semantic interoperability of metadata,”AI Magazine, vol. 24, no. 3, pp. 75–92, 2003

    work page 2003

  6. [6]

    The semantic web,

    T. Berners-Lee, J. Hendler, and O. Lassila, “The semantic web,”Scien- tific American, vol. 284, no. 5, pp. 34–43, 2001

    work page 2001

  7. [7]

    Graph neural networks in vision-language image understanding: a survey: Graph neural networks in vision-language image understanding: a survey,

    H. Senior, G. Slabaugh, S. Yuan, and L. Rossi, “Graph neural networks in vision-language image understanding: a survey: Graph neural networks in vision-language image understanding: a survey,”Vis. Comput., vol. 41, no. 1, p. 491–516, Mar. 2024. [Online]. Available: https://doi.org/10.1007/s00371-024-03343-0

    work page doi:10.1007/s00371-024-03343-0 2024

  8. [8]

    Yolov11: An overview of the key architectural enhancements,

    R. Khanam and M. Hussain, “Yolov11: An overview of the key architectural enhancements,”arXiv preprint arXiv:2410.17725, 2024. [Online]. Available: https://arxiv.org/abs/2410.17725

    arXiv 2024

Show all 47 references
  1. [9]

    Detectron2,

    Y . Wu, A. Kirillov, F. Massa, W.-Y . Lo, and R. Girshick, “Detectron2,” https://github.com/facebookresearch/detectron2, 2019

    2019

  2. [10]

    Deep high-resolution representation learning for visual recognition,

    J. Wang, K. Sun, T. Cheng, B. Jiang, C. Deng, Y . Zhao, D. Liu, Y . Mu, M. Tan, X. Wang, W. Liu, and B. Xiao, “Deep high-resolution representation learning for visual recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. [Online]. Available: https:...

    2020 arXiv

  3. [11]

    On the opportunities and risks of foundation models,

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskillet al., “On the opportunities and risks of foundation models,”arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  4. [12]

    A review on deep learning techniques applied to semantic segmentation,

    A. Garcia-Garcia, S. Orts-Escolano, S. Oprea, V . Villena-Martinez, and J. Garcia-Rodriguez, “A review on deep learning techniques applied to semantic segmentation,”CoRR, vol. abs/1704.06857, 2017. [Online]. Available: http://arxiv.org/abs/1704.06857

    2017 arXiv

  5. [13]

    Gold panning in vocabulary: An adaptive method for vocabulary expansion of domain-specific llms,

    C. Liu, S. Wang, L. Qing, K. Kuang, Y . Kang, C. Sun, and F. Wu, “Gold panning in vocabulary: An adaptive method for vocabulary expansion of domain-specific llms,” 2024. [Online]. Available: https://arxiv.org/abs/2410.01188

    2024 arXiv

  6. [14]

    Knowledge graph- enhanced large language models via path selection,

    H. Liu, S. Wang, Y . Zhu, Y . Dong, and J. Li, “Knowledge graph- enhanced large language models via path selection,” 2024. [Online]. Available: https://arxiv.org/abs/2406.13862

    2024 arXiv

  7. [15]

    Europeana data model primer,

    A. Isaac and B. Haslhofer, “Europeana data model primer,”Europeana Foundation, 2009. [Online]. Available: https://pro.europeana.eu/ files/Europeana Professional/Share your data/Technical requirements/ EDM Documentation/EDM Primer 130714.pdf

    2009

  8. [16]

    Rethinking atrous convolution for semantic image segmentation,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A. L. Yuille, “Rethinking atrous convolution for semantic image segmentation,” arXiv preprint arXiv:1706.05587, 2017. [Online]. Available: https: //arxiv.org/abs/1706.05587

    2017 arXiv

  9. [18]

    Knowledge graphs meet multi-modal learning: A comprehensive survey,

    Z. Chen, Y . Zhang, Y . Fang, Y . Geng, L. Guo, X. Chen, Q. Li, W. Zhang, J. Chen, Y . Zhu, J. Li, X. Liu, J. Z. Pan, N. Zhang, and H. Chen, “Knowledge graphs meet multi-modal learning: A comprehensive survey,” 2024. [Online]. Available: https://arxiv.org/abs/2402.05391

    2024 arXiv

  10. [19]

    An effective approach to text detection and recognition in degraded historical documents,

    P. Maldonado-Quispe and H. Pedrini, “An effective approach to text detection and recognition in degraded historical documents,” inProgress in Pattern Recognition, Image Analysis, Computer Vision, and Applica- tions, R. Hern ´andez-Garc´ıa, R. J. Barrientos, and S. A. Velastin,...

    2025

  11. [20]

    Wikidata: A free collaborative knowl- edgebase,

    D. Vrande ˇci´c and M. Kr ¨otzsch, “Wikidata: A free collaborative knowl- edgebase,”Communications of the ACM, vol. 57, no. 10, pp. 78–85, 2014

    2014

  12. [21]

    The fair guiding principles for scientific data management,

    M. D. Wilkinsonet al., “The fair guiding principles for scientific data management,”Scientific Data, vol. 3, p. 160018, 2016

    2016

  13. [22]

    Jagiellonian digital library,

    Jagiellonian University, “Jagiellonian digital library,” https://jbc.bj.uj. edu.pl/dlibra, 2023, [Online; accessed 20-June-2023]

    2023

  14. [23]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    M. Tan and Q. V . Le, “Efficientnet: Rethinking model scaling for convolutional neural networks,” 2020. [Online]. Available: https: //arxiv.org/abs/1905.11946

    2020 arXiv

  15. [24]

    Exploring entity recognition and disambiguation for cultural heritage collections,

    S. van Hooland, M. De Wilde, R. Verborgh, T. Steiner, and R. Van de Walle, “Exploring entity recognition and disambiguation for cultural heritage collections,”Digital Scholarship in the Humanities, vol. 30, no. 2, pp. 262–279, Jun. 2015. [Online]. Available: http://freeyourmet...

    2015

  16. [25]

    Deep residual learning for image recognition,

    K. Heet al., “Deep residual learning for image recognition,”CVPR, pp. 770–778, 2016

    2016

  17. [26]

    Linked Data - Design Issues,

    Tim Berners-Lee, “Linked Data - Design Issues,” 2006. [Online]. Available: https://www.w3.org/DesignIssues/LinkedData.html

    2006

  18. [27]

    Think culture: Europeana.eu from concept to construction,

    J. Purday, “Think culture: Europeana.eu from concept to construction,” The Electronic Library, vol. 27, no. 6, pp. 919–937, 2009

    2009

  19. [28]

    Knowledge graph visualization interface for digital heritage collections: Design issues and recommendations,

    C. S. Khoo, E. A. Tan, S.-G. Ng, C.-F. Chan, M. Stanley-Baker, and W.-N. Cheng, “Knowledge graph visualization interface for digital heritage collections: Design issues and recommendations,”Information Technology and Libraries, vol. 43, no. 1, Mar. 2024. [Online]. Available: h...

    2024

  20. [29]

    Ollama: Run Large Language Models Locally,

    Ollama Team, “Ollama: Run Large Language Models Locally,” 2025. [Online]. Available: https://github.com/ollama/ollama

    2025

  21. [30]

    DeepSeek v3: Advanced AI Language Model,

    DeepSeek AI Team, “DeepSeek v3: Advanced AI Language Model,”

  22. [31]

    Knowledge graphs for enhancing large language models in entity disambiguation,

    G. Pons, B. Bilalli, and A. Queralt, “Knowledge graphs for enhancing large language models in entity disambiguation,” inThe Semantic Web – ISWC 2024, G. Demartini, K. Hose, M. Acosta, M. Palmonari, G. Cheng, H. Skaf-Molli, N. Ferranti, D. Hern ´andez, and A. Hogan, Eds. Cham: ...

    2024

  23. [32]

    Better together: Enhancing generative knowledge graph completion with language models and neighborhood information,

    A. Chepurova, A. Bulatov, Y . Kuratov, and M. Burtsev, “Better together: Enhancing generative knowledge graph completion with language models and neighborhood information,” inFindings of the Association for Computational Linguistics: EMNLP 2023, H. Bouamor, J. Pino, and K. Bal...

    2023

  24. [33]

    Enhancing entity alignment between wikidata and artgraph using llms,

    A. S. Lippolis, A. Klironomos, D. F. Milon-Flores, H. Zheng, A. Jouglar, E. Norouzi, and A. Hogan, “Enhancing entity alignment between wikidata and artgraph using llms,” inInternational Joint Workshop on Semantic Web and Ontology Design for Cultural Heritage,

  25. [34]

    Using large language models for knowledge engineering (llmke): A case study on wikidata,

    B. Zhang, I. Reklos, N. Jain, A. Pe ˜nuela, and E. Simperl, “Using large language models for knowledge engineering (llmke): A case study on wikidata,”CEUR Workshop Proceedings, vol. 3577, Jan. 2023

    2023

  26. [35]

    Knowledge Graph Construction Using Large Language Models,

    Y . Lairgi, L. Moncla, R. Cazabet, K. Benabdeslem, and P. Cl ´eau, “Knowledge Graph Construction Using Large Language Models,” in Journ´ee nationale sur la fouille de textes, Lyon, France, Jun. 2024. [Online]. Available: https://hal.science/hal-04607294

    2024

  27. [36]

    Available: https://api.semanticscholar.org/CorpusID: 263913092

    [Online]. Available: https://api.semanticscholar.org/CorpusID: 263913092

  28. [37]

    Separate and locate: Rethink the text in text-based visual question answering,

    C. Fang, J. Li, L. Li, C. Ma, and D. Hu, “Separate and locate: Rethink the text in text-based visual question answering,” inProceedings of the 31st ACM International Conference on Multimedia, ser. MM ’23. New York, NY , USA: Association for Computing Machinery, 2023, p. 4378–4...

    2023

  29. [38]

    Locate then generate: bridging vision and language with bounding box for scene-text vqa,

    Y . Zhu, Z. Liu, Y . Liang, X. Li, H. Liu, C. Bao, and L. Xu, “Locate then generate: bridging vision and language with bounding box for scene-text vqa,” inProceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative App...

    2023 doi

  30. [39]

    Accelerating knowledge graph and ontology engineering with large language models,

    C. Shimizu and P. Hitzler, “Accelerating knowledge graph and ontology engineering with large language models,”Journal of Web Semantics, p. 100862, 2025. [Online]. Available: https://www.sciencedirect.com/ science/article/pii/S1570826825000022

    2025

  31. [40]

    Knowledge graph- enhanced large language models via path selection,

    H. Liu, S. Wang, Y . Zhu, Y . Dong, and J. Li, “Knowledge graph- enhanced large language models via path selection,” inFindings of the Association for Computational Linguistics: ACL 2024, L.-W. Ku, A. Martins, and V . Srikumar, Eds. Bangkok, Thailand: Association for Computati...

    2024

  32. [41]

    Handwriting recognition in historical documents with multimodal llm,

    L. Li, “Handwriting recognition in historical documents with multimodal llm,” 2024. [Online]. Available: https://arxiv.org/abs/2410.24034

    2024 arXiv

  33. [42]

    Publishing the british national bibliography as linked open data,

    B. Library, “Publishing the british national bibliography as linked open data,”British Library, 2011. [Online]. Available: https://www.bl.uk/ bibliographic/pdfs/british national bibliography linked open data.pdf

    2011

  34. [43]

    Enabling llms to generate fine-grained citations in long- context qa,

    J. Zhanget al., “Enabling llms to generate fine-grained citations in long- context qa,”arXiv preprint arXiv:2409.02897, 2024

    2024 arXiv

  35. [44]

    Refining input guardrails: Enhancing llm-as- a-judge efficiency through chain-of-thought fine-tuning and alignment,

    M. Kazemi Radet al., “Refining input guardrails: Enhancing llm-as- a-judge efficiency through chain-of-thought fine-tuning and alignment,” arXiv preprint arXiv:2501.13080, 2025

    2025 arXiv

  36. [45]

    Enabling large language models to generate text with citations,

    T. Gao, H. Yen, J. Yu, and D. Chen, “Enabling large language models to generate text with citations,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 6465– 6488

    2023

  37. [48]

    Ruler: Improving llm controllability by rule-based data recycling,

    M. Liet al., “Ruler: Improving llm controllability by rule-based data recycling,”arXiv preprint arXiv:2406.15938, 2024

    2024 arXiv

  38. [2023]

    Available: https://arxiv.org/abs/2303.08774

    [Online]. Available: https://arxiv.org/abs/2303.08774

    arXiv

  39. [2025]

    Available: https://github.com/deepseek-ai/DeepSeek-V3

    [Online]. Available: https://github.com/deepseek-ai/DeepSeek-V3

Pith tools

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