Pith. sign in

REVIEW 3 major objections 4 minor 40 references

CORE-KG: An LLM-Driven Knowledge Graph Construction Framework for Human Smuggling Networks

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CORE-KG, a two-step LLM framework, builds cleaner knowledge graphs of human smuggling networks from court documents, cutting duplicate nodes by 33.28% and legal noise by 38.37% compared to a GraphRAG baseline.

desk verdict Useful applied paper with a real evaluation gap: the two headline metrics are hard to interpret because they measure whether the prompt followed its own instructions, not graph quality. read the letter →

arxiv 2506.21607 v1 pith:5ELI4VAH submitted 2025-06-20 cs.CL cs.AIcs.LG

Dipak Meher , Carlotta Domeniconi , Guadalupe Correa-Cabrera This is my paper
classification cs.CLcs.AIcs.LG
keywords knowledgegraphconstructioncoreferenceresolutionLLMpromptinghumansmugglingnetworksRAGlegaltextanalysisentityextractionnodeduplication
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 sets out to establish that knowledge graphs of human smuggling networks can be built from unstructured legal documents with far less fragmentation and clutter than standard LLM pipelines produce. It proposes CORE-KG, which first resolves coreferences one entity type at a time—collapsing aliases, abbreviations, and role-based mentions into single canonical names—and then extracts entities and relationships with a prompt that sequences entity types, defines each type explicitly, and instructs the model to drop government-related legal boilerplate. Across twenty U.S. court cases, this reduces node duplication by 33.28% and legal noise by 38.37% relative to a minimally adapted GraphRAG baseline. A sympathetic reader would care because cleaner graphs make downstream analysis of smuggling networks, such as identifying key actors and routes, more reliable.

What carries the argument

The load-bearing mechanism is the type-aware sequential coreference resolution module, applied in a fixed order to seven entity types—Person, Location, Routes, Organization, Means of Transportation, Means of Communication, and Smuggled Items—with each pass rewriting the input text to a canonical form before the next type is processed. A tailored extraction prompt then guides entity and relationship extraction inside an adapted GraphRAG pipeline, adding three controls: sequential extraction of entity types to reduce attention diffusion, in-prompt type definitions to counter overgeneralization bias, and an explicit filtering step that removes government-related legal boilerplate. These two components together are what the authors credit for the reported reductions in duplication and noise.

What would settle it

A blinded re-evaluation in which annotators do not know which pipeline produced each graph, and in which the noise metric excludes exactly the entities the prompt orders removed, would settle the claim: if CORE-KG and the GraphRAG baseline show no significant difference in duplication or noise under those conditions, the reported improvements are artifacts of the measurement. Likewise, if re-running the pipeline with a different instruction-tuned LLM erases the gap, the gains would be attributable to the specific model rather than to the framework.

Watch

Extended reading notes

Core claim

The central claim is that a modular LLM pipeline that separates type-aware coreference resolution from extraction, and performs both in a structured, prompt-guided way, yields knowledge graphs that are both less redundant and less dominated by irrelevant legal entities than the unmodified GraphRAG pipeline. In the paper's own account, the coreference module unifies surface variants like 'Y.', 'A.Y.', and 'the defendant' into a single canonical node, and the extraction prompt's explicit filtering instruction removes procedural terms such as 'court', 'appeal', and 'judicial proceedings' during extraction itself. On the 20-case evaluation, the node duplication rate falls from 30.38% to 20.27% and the legal noise rate from 27.41% to 16.89%.

Load-bearing premise

The manual and fuzzy duplicate and noise counts in Section 5.1 are treated as unbiased measures of graph quality, even though the noise definition counts legal boilerplate entities that the CORE-KG prompt explicitly instructs the model to filter out in Section 3.3.1.

Editorial extensions

If this is right

  • If CORE-KG is correct, applying the same two-step recipe to other dense legal or investigative corpora should yield proportionally fewer duplicate nodes and fewer irrelevant procedural entities than a plain GraphRAG pipeline.
  • The sequential, type-wise coreference design implies that separating entity types reduces cross-type interference, so mixing all seven types in a single resolution pass should produce more fragmented graphs on similar texts.
  • The in-prompt filtering instruction implies that a graph's noise level is partly a prompt-design choice; downstream criminal-network analyses should prefer graphs built with such filters to avoid domination by courts and appeals.
  • Because the extraction prompt enumerates exactly seven entity types, the framework is directly adaptable to other fixed schemas of entities relevant to a given investigation.

Reading between the lines

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

  • The paper compares only against a GraphRAG baseline and does not validate against an annotated ground-truth graph; a direct comparison with human-constructed graphs would be a stronger test of whether the reported reductions correspond to true quality gains rather than relative differences.
  • Since the noise metric counts exactly the legal boilerplate entities that the CORE-KG prompt orders the model to exclude, the 38.37% noise reduction may partly reflect prompt obedience; a blinded evaluation or a noise definition that excludes prompt-mandated omissions would isolate the framework's independent effect.
  • Because only one LLM (LLaMA 3.3 70B at temperature zero) is used, the generality of the gains across different instruction-tuned models remains untested; re-running with a second model would probe the methodology rather than the model.
  • A testable extension the authors do not run is a downstream task evaluation—for example, community detection or link prediction on the resulting graphs—to check whether the cleaner structure actually improves analysis of smuggling networks.
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 CORE-KG, a modular LLM-based framework for constructing knowledge graphs from legal case documents about human smuggling networks. The pipeline first performs entity-type-aware coreference resolution with sequential structured prompts, then adapts GraphRAG's entity/relation extraction with domain-specific instructions, including sequential type extraction, explicit type definitions, and filtering of government/legal-boilerplate entities. The authors evaluate on 20 U.S. federal and state court cases and report that CORE-KG reduces node duplication by 33.28% and legal noise by 38.37% relative to a lightly adapted GraphRAG baseline, concluding that the framework produces cleaner and more coherent graphs for criminal network analysis.

Significance. If the quantitative claims were validated, CORE-KG would be a useful contribution to LLM-based knowledge graph construction in a challenging, low-resource legal domain. The paper's strengths include a reproducible open-source implementation, a deterministic decoding setup (temperature zero), and qualitative examples that show plausible context-sensitive coreference behavior (e.g., not merging distinct concepts in Case 6, correctly grouping 'relatives of the aliens' in Case 13). However, the two headline metrics are confounded with the method's own design choices: the noise reduction is largely predetermined by the prompt's filtering instruction, and the duplication metric cannot detect over-merging errors. The absence of ground truth, error bars, and inter-annotator agreement means that the central empirical claim is not yet established. The contribution is best viewed as a promising system description pending a more rigorous evaluation.

major comments (3)
  1. [§5.1 Noise Detection vs. §3.3.1] The noise rate is defined over legal/government entities such as 'Court,' 'Appeal,' and 'Judicial Proceedings,' which are exactly the categories that the CORE-KG extraction prompt explicitly instructs the model to filter out (Appendix Figure 3: 'Do not extract entities corresponding to governmental organizations...'). The GraphRAG baseline is not given this instruction. The reported 38.37% noise reduction therefore largely measures whether CORE-KG follows its own filtering rule, not whether the resulting graphs are more analytically useful. An independent evaluation—for example, human relevance judgments on a sample of nodes, or measuring noise for entity types not mentioned in the prompt—is needed to support RQ2.
  2. [§5.1 Duplicate Node Detection] The node duplication metric only penalizes under-merging: it counts cases where multiple nodes should have been merged, but it has no way to detect over-merging, where distinct real-world entities are incorrectly fused during coreference resolution. Over-merging reduces the node count and mechanically lowers the duplication rate, so the 33.28% reduction could partly reflect erroneous consolidation that the manual/fuzzy review cannot see. The manual review is described as correcting false positives in fuzzy clusters but is not blinded and has no reported inter-annotator agreement; no precision or recall against a gold standard is given. The authors should report a false-merge rate on a sample of resolved coreference chains, or compare against a human-annotated entity resolution set.
  3. [§4.1 and §5.1] The evaluation is based on only 20 cases with no annotated ground-truth graphs, no confidence intervals, and no significance testing. The paper claims 'consistent performance improvements' but Appendix Figures 1 and 2 show cases where the baseline outperforms CORE-KG (e.g., Cases 6 and 7 for noise). Aggregated averages without error bars or paired tests do not support the strength of the claims in the abstract. A paired bootstrap or Wilcoxon signed-rank test, plus per-case scatter plots with confidence intervals, would be a minimal addition.
minor comments (4)
  1. [§5.1] The section heading 'Impact of Coreference Resolution on Node' is incomplete; it should finish with 'Duplication' or similar.
  2. [Appendix C] The sentence 'Other coreferences have also been resolved and resultant into clean and more coherent graph' contains grammatical errors ('resultant into') and should be rewritten.
  3. [§4.2] The paper states temperature is set to zero, but this does not eliminate all sampling nondeterminism; this is worth acknowledging, and ideally the authors would report whether runs were repeated and stable.
  4. [§5.1.2] The error analysis mentions missed coreference cases (e.g., Case 15 and Case 10), but no quantitative rate of missed resolutions is reported; adding such a rate would help readers judge the trade-off between precision and recall in coreference merging.

Circularity Check

1 steps flagged · score 6.0 of 10

Noise-reduction metric is predetermined by CORE-KG's own filtering instruction; the duplication metric is independent in design but the evaluation lacks a gold standard.

  1. self definitional [Section 3.3.1 (Prompt Tuning) and Section 5.1 (Noise Detection); Appendix B Figure 3]
    "we introduced an explicit filtering instruction within the prompt. After completing the entity and relationship extraction steps, the LLM is guided to identify all government-related nodes based on predefined criteria and remove them before producing the final output. // We quantify noise using the noise rate, defined as the number of non-informative nodes divided by the total number of nodes in the graph."

    The noise metric is defined by counting legal/government entities such as Court, Appeal, and Judicial Proceedings, and the CORE-KG extraction prompt explicitly instructs the model to exclude exactly those entities. The GraphRAG baseline receives no such exclusion instruction. Thus the reported 38.37% relative noise reduction largely confirms that the model follows its own prompt rule; it is not an independent measurement of graph quality or analytical relevance. The input design defines the evaluation quantity by construction, making this headline result predetermined up to LLM instruction-following.

full rationale

The duplication-rate result is not circular: CORE-KG's coreference-resolution prompt is tested against a fuzzy-matching plus expert-review duplicate count, and the 33.28% relative reduction is an empirical outcome of the method rather than a quantity fixed by the method's definition. That said, the evaluation has correctness risks not captured by circularity: there is no gold-standard precision/recall check, and the duplicate count cannot detect over-merging, which would lower the reported duplication rate while corrupting the graph. The noise-rate result, however, is largely circular by construction: the set of entities counted as noise is the same set the extraction prompt explicitly instructs the model to remove, so the 38.37% reduction mainly verifies prompt compliance. No load-bearing self-citations were found; the cited GraphRAG baseline is external, and the author's own prior work appears only as a generic LLM capability citation. Because the paper's central 'cleaner and more coherent graphs' claim rests on two headline metrics, and one of them reduces to the method's own instruction, the paper has partial circularity rather than full circularity.

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

The central claims rest on the reliability of LLM coreference resolution and on the validity of the manual evaluation protocol. No new entities are invented. The free parameters are evaluation and preprocessing choices that directly affect the measured improvements.

free parameters (4)
  • fuzzy match similarity threshold = 75% partial_ratio
    Used in duplicate node detection; controls which mention pairs are treated as duplicates and directly affects the reported duplication rate (Section 5.1).
  • chunk size = 300 tokens
    GraphRAG input chunk size; affects extraction context and entity merging (Section 4.2).
  • entity type set = 7 types: Person, Location, Organization, Route, Means of Transportation, Means of Communication, Smuggled Items
    Chosen with a subject matter expert; if these types are not the right ontology, the graph quality claims do not transfer (Section 3.2.1).
  • noise definition = legal boilerplate such as court, appeal, judicial proceedings
    The noise metric is defined over categories that the CORE-KG prompt explicitly filters out, making the noise reduction partly by construction (Section 5.1 vs 3.3.1).
assumptions (4)
  • domain assumption LLMs can perform coreference resolution on legal texts reliably when prompted per entity type.
    The entire method depends on LLaMA 3.3 70B resolving mentions such as 'Y.' to 'A.Y.' correctly; no ground-truth evaluation of the coreference module itself is provided (Section 3.2).
  • domain assumption The Opinion section of each case contains the factual narrative sufficient for KG construction.
    The pipeline discards all other sections; if key smuggling entities appear outside Opinion, graphs are incomplete (Section 4.1).
  • domain assumption Duplicate clusters found by fuzzy matching plus manual review correspond to true coreferent real-world entities.
    The duplication metric is computed from this clustering; errors directly change the reported improvement (Section 5.1).
  • domain assumption The GraphRAG baseline with minimal adaptation is a fair comparison point.
    All improvements are relative to this baseline; if the baseline prompt is under-optimized, the gains overstate CORE-KG's advantage (Section 4.3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of CORE-KG: An LLM-Driven Knowledge Graph Construction Framework for Human Smuggling Networks." pith.science (2026). https://pith.science/paper/5ELI4VAH

@misc{pith2026250621607,
  author       = {Pith},
  title        = {Pith review of: CORE-KG: An LLM-Driven Knowledge Graph Construction Framework for Human Smuggling Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5ELI4VAH}},
  note         = {Machine review of arXiv:2506.21607}
}
read the original abstract

Human smuggling networks are increasingly adaptive and difficult to analyze. Legal case documents offer valuable insights but are unstructured, lexically dense, and filled with ambiguous or shifting references-posing challenges for automated knowledge graph (KG) construction. Existing KG methods often rely on static templates and lack coreference resolution, while recent LLM-based approaches frequently produce noisy, fragmented graphs due to hallucinations, and duplicate nodes caused by a lack of guided extraction. We propose CORE-KG, a modular framework for building interpretable KGs from legal texts. It uses a two-step pipeline: (1) type-aware coreference resolution via sequential, structured LLM prompts, and (2) entity and relationship extraction using domain-guided instructions, built on an adapted GraphRAG framework. CORE-KG reduces node duplication by 33.28%, and legal noise by 38.37% compared to a GraphRAG-based baseline-resulting in cleaner and more coherent graph structures. These improvements make CORE-KG a strong foundation for analyzing complex criminal networks.

Figures

Figures reproduced from arXiv: 2506.21607 by the authors.

Figure 3
Figure 3. In the same legal case, CORE-KG extracted no legal noise entities, yielding a clean and focused graph. Quantitatively, CORE-KG re￾duces the overall legal noise rate from 27.41% to 16.89%, an absolute improvement of 10.52% points ( see [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 1
Figure 1. Node duplication rate comparison between baseline [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Noise reduction rate comparison between baseline [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Prompt used for entity and relationship extraction [PITH_FULL_IMAGE:figures/full_fig_p010_3.png]
Figure 6
Figure 6. Figure 6: Prompt used for resolving coreferences related to [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 4
Figure 4. Figure 4: Knowledge graph generated by CORE-KG for a representative legal case. The graph demonstrates resolved coreference, [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Baseline knowledge graph generated using GraphRAG for a representative legal case. The graph contains several [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 27 canonical work pages

  1. [1]

    Sahar Abdelnabi, Aideen Fay, Giovanni Cherubin, Ahmed Salem, Mario Fritz, and Andrew Paverd. 2024. Are you still on track!? Catching LLM Task Drift with Activations.arXiv preprint arXiv:2406.00799(2024)

    arXiv 2024

  2. [2]

    Sahar Abdelnabi, Aideen Fay, Giovanni Cherubin, Ahmed Salem, Mario Fritz, and Andrew Paverd. 2025. Get My Drift? Catching LLM Task Drift with Activation Deltas. In2025 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML). IEEE, 43–67

    work page 2025

  3. [3]

    Shengnan An, Zexiong Ma, Zeqi Lin, Nanning Zheng, Jian-Guang Lou, and Weizhu Chen. 2024. Make your llm fully utilize the context.Advances in Neural Information Processing Systems37 (2024), 62160–62188

    2024

  4. [4]

    Tiffany J Callahan, Ignacio J Tripodi, Adrianne L Stefanski, Luca Cappelletti, Sanya B Taneja, Jordan M Wyrwa, Elena Casiraghi, Nicolas A Matentzoglu, Justin Reese, Jonathan C Silverstein, et al . 2024. An open source knowledge graph ecosystem for the life sciences.Scientific Data11, 1 (2024), 363

    work page 2024

  5. [5]

    Andrew Carlson, Justin Betteridge, Richard C Wang, Estevam R Hruschka Jr, and Tom M Mitchell. 2010. Coupled semi-supervised learning for information extraction. InProceedings of the third ACM international conference on Web search and data mining. 101–110

    work page 2010

  6. [6]

    Sara Carrasco-Granger, Rut Bermejo-Casado, and Isabel Bazaga Fernández. 2025. Scapegoating Human Smugglers: How Migrants’ Accounts Challenge the EU’s Dominant Narrative.Journal of Immigrant & Refugee Studies(2025), 1–13

    work page 2025

  7. [7]

    Penghe Chen, Yu Lu, Vincent W Zheng, Xiyang Chen, and Boda Yang. 2018. Knowedu: A system to construct knowledge graph for education.Ieee Access6 (2018), 31553–31563

    work page 2018

  8. [8]

    Sunhao Dai, Chen Xu, Shicheng Xu, Liang Pang, Zhenhua Dong, and Jun Xu

Show all 40 references
  1. [9]

    Mridusmita Das and Apurbalal Senapati. 2024. Co-reference Resolution in Prompt Engineering.Procedia Computer Science244 (2024), 194–201

    2024

  2. [10]

    Yiran Ding, Li Lyna Zhang, Chengruidong Zhang, Yuanyuan Xu, Ning Shang, Jiahang Xu, Fan Yang, and Mao Yang. 2024. Longrope: Extending llm context window beyond 2 million tokens.arXiv preprint arXiv:2402.13753(2024). CORE-KG: An LLM-Driven Knowledge Graph Construction Framework...

    2024 arXiv

  3. [11]

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. 2024. From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130(2024)

    2024 arXiv

  4. [12]

    Weili Fang, Ling Ma, Peter ED Love, Hanbin Luo, Lieyun Ding, and AO Zhou

  5. [13]

    Yujian Gan, Juntao Yu, and Massimo Poesio. 2024. Assessing the capabilities of large language models in coreference: An evaluation. InJoint 30th International Conference on Computational Linguistics and 14th International Conference on Language Resources and Evaluation, LREC-C...

    2024

  6. [14]

    Alex Havrilla, Sharath Raparthy, Christoforus Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, and Roberta Raileanu. 2024. Glore: When, where, and how to improve llm reasoning via global and local refinements.arXiv preprint arXiv:2402.10963(2024)

    2024 arXiv

  7. [15]

    Elwin Huaman, Elias Kärle, and Dieter Fensel. 2020. Duplication detection in knowledge graphs: Literature and tools.arXiv preprint arXiv:2004.08257(2020)

    2020 arXiv

  8. [16]

    Donghong Ji, Jun Gao, Hao Fei, Chong Teng, and Yafeng Ren. 2020. A deep neural network model for speakers coreference resolution in legal texts.Information Processing & Management57, 6 (2020), 102365

    2020

  9. [17]

    Nanda Kambhatla. 2004. Combining lexical, syntactic, and semantic features with maximum entropy models for information extraction. InProceedings of the ACL interactive poster and demonstration sessions. 178–181

    2004

  10. [18]

    Vamsi Krishna Kommineni, Birgitta König-Ries, and Sheeba Samuel. 2024. From human experts to machines: An LLM supported approach to ontology and knowl- edge graph construction.arXiv preprint arXiv:2403.08345(2024)

    2024 arXiv

  11. [19]

    Kusum Lata, Pardeep Singh, and Kamlesh Dutta. 2022. Mention detection in coreference resolution: survey.Applied Intelligence52, 9 (2022), 9816–9860

    2022

  12. [20]

    Nghia T Le and Alan Ritter. 2023. Are Large Language Models Robust Coreference Resolvers?arXiv preprint arXiv:2305.14489(2023)

    2023 arXiv

  13. [21]

    Xiao Ling and Daniel Weld. 2012. Fine-grained entity recognition. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 26. 94–100

    2012

  14. [22]

    Kebin Liu, Fang Li, Lei Liu, and Ying Han. 2007. Implementation of a kernel-based Chinese relation extraction system.Jisuanji Yanjiu yu Fazhan(Computer Research and Development)44, 8 (2007), 1406–1411

    2007

  15. [23]

    Ruicheng Liu, Rui Mao, Anh Tuan Luu, and Erik Cambria. 2023. A brief survey on recent advances in coreference resolution.Artificial Intelligence Review56, 12 (2023), 14439–14481

    2023

  16. [24]

    Xiaohua Liu, Shaodian Zhang, Furu Wei, and Ming Zhou. 2011. Recognizing named entities in tweets. InProceedings of the 49th annual meeting of the associa- tion for computational linguistics: human language technologies. 359–367

    2011

  17. [25]

    Svitlana Mazepa, Victoria Vysotska, Danylo Ivanchyshyn, Lyubomyr Chyrun, Vadim Schuchmann, and Yuriy Ryshkovets. 2022. Relationships Knowledge Graphs Construction Between Evidence Based on Crime Reports. In2022 IEEE 17th International Conference on Computer Sciences and Inform...

    2022

  18. [26]

    Dipak Falgun Meher, Ajay Krishna Vajjala, and David S Rosenblum. [n. d.]. Un- derstanding User Behavior in Cross-Domain Recommendation: An LLM-Based Approach. ([n. d.])

  19. [27]

    Uwe Peters and Benjamin Chin-Yee. 2025. Generalization bias in large language model summarization of scientific research.Royal Society Open Science12, 4 (2025), 241776

    2025

  20. [28]

    Sergiy Pogorilyy and Artem Kramov. 2019. Coreference resolution method using a convolutional neural network. In2019 IEEE International Conference on Advanced Trends in Information Theory (ATIT). IEEE, 397–401

    2019

  21. [29]

    Yong Shi, Wenlu An, Jiayu Xue, and Yi Qu. 2022. A knowledge graph constructed for job-related crimes.Procedia Computer Science199 (2022), 540–547

    2022

  22. [30]

    Peng Sun, Xuezhen Yang, Xiaobing Zhao, and Zhijuan Wang. 2018. An overview of named entity recognition. In2018 International Conference on Asian Language Processing (IALP). IEEE, 273–278

    2018

  23. [31]

    Hieu-Dai Tran, Duc-Vu Nguyen, and Ngan Luu-Thuy Nguyen. 2025. Coreference Resolution for Vietnamese Narrative Texts.arXiv preprint arXiv:2504.19606 (2025)

    2025 arXiv

  24. [32]

    Julio Vizcarra, Shuichiro Haruta, and Mori Kurokawa. 2024. Representing the interaction between users and products via llm-assisted knowledge graph con- struction. In2024 IEEE 18th International Conference on Semantic Computing (ICSC). IEEE, 231–232

    2024

  25. [33]

    Boshi Wang, Xiang Yue, and Huan Sun. 2023. Can ChatGPT defend its belief in truth? evaluating LLM reasoning via debate.arXiv preprint arXiv:2305.13160 (2023)

    2023 arXiv

  26. [34]

    Tai Wang and Huan Li. 2020. Coreference resolution improves educational knowl- edge graph construction. In2020 IEEE International Conference on Knowledge Graph (ICKG). IEEE, 629–634

    2020

  27. [35]

    Jheng-Long Wu and Wei-Yun Ma. 2017. A deep learning framework for coref- erence resolution based on convolutional neural network. In2017 IEEE 11th International Conference on Semantic Computing (ICSC). IEEE, 61–64

    2017

  28. [36]

    Bowen Zhang and Harold Soh. 2024. Extract, define, canonicalize: An llm-based framework for knowledge graph construction.arXiv preprint arXiv:2404.03868 (2024)

    2024 arXiv

  29. [37]

    Lingfeng Zhong, Jia Wu, Qian Li, Hao Peng, and Xindong Wu. 2023. A compre- hensive survey on automatic knowledge graph construction.Comput. Surveys 56, 4 (2023), 1–62

    2023

  30. [38]

    Agent R

    Wenxuan Zhou, Sheng Zhang, Yu Gu, Muhao Chen, and Hoifung Poon. 2023. Universalner: Targeted distillation from large language models for open named entity recognition.arXiv preprint arXiv:2308.03279(2023). A Per-Case Duplication and Noise Metrics Figure 1: Node duplication rat...

    2023 arXiv

  31. [2020]

    Knowledge graph for identifying hazards on construction sites: Integrating computer vision with ontology.Automation in Construction119 (2020), 103310

    2020

  32. [2024]

    InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

    Bias and unfairness in information retrieval systems: New challenges in the llm era. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6437–6447

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.