Pith. sign in

REVIEW 4 major objections 3 minor 36 references

The paper claims that live retrieval of a small ontology slice, injected into an LLM extraction prompt, keeps a 9B local model aligned to a formal schema while cutting prompt overhead by about 94 percent, and that layered zero-inference ded

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 00:44 UTC pith:JOK7SFW6

load-bearing objection Solid engineering report, but the headline numbers are development-corpus fits, not predictions. the 4 major comments →

arxiv 2607.28662 v1 pith:JOK7SFW6 submitted 2026-07-22 cs.AI

An Ontology-Guided, Deduplication-Aware Extraction Layer for Knowledge Graph Construction from Heterogeneous Documents

Vaibhav Dangaich , Kevin Lewis , Kundeshwar Pundalik This is my paper
classification cs.AI
keywords ontology-guided extractionknowledge graph constructionentity resolutiondeduplicationretrieval-augmented generationper-page OCR classificationLLM extractionhierarchical ontology retrieval
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The authors aim to show that a production extraction layer can turn a live, heterogeneous document stream into a validated, ontology-aligned knowledge graph using a modest locally hosted LLM, provided the model is steered by live graph retrieval and repaired by layered deduplication. Unguided LLM extraction fractures type vocabularies and duplicates entities; the system constrains and repairs the model at every stage. If correct, ontology-guided retrieval plus six rule-based deduplication algorithms plus an embedding-based resolution stage with an unoverridable conflict guard is sufficient to make a 9B model production-viable for schema-governed graph construction. The headline numbers are measured on two synthetic development documents, with all thresholds tuned on those documents; the paper frames them as directional engineering evidence, not statistically validated effect sizes.

Core claim

Ontology grounding at extraction time, implemented as live vector retrieval of relevant classes and predicates from a graph database and injection of only those into the prompt, replaces static domain catalog slices and cuts catalog prompt overhead by roughly 94 percent. Combined with a five-stage refinement pipeline (deterministic cleaning, cross-chunk merging, quality-gated relationship second pass, six rule-based deduplication algorithms, and an embedding-based resolution engine whose hard-conflict guard no similarity score can override), search recall rises from roughly 70 to 95 percent while maintaining zero false merges. The paper also demonstrates that per-page OCR classification rout

What carries the argument

The load-bearing mechanism is ontology-guided retrieval: a content sample is split into overlapping windows, embedded, and used to query a vector index over ontology class and predicate definitions; candidates are scored by cosine similarity penalized by PageRank centrality, thresholded at 0.72, and trimmed to a token budget, then injected into the extraction prompt. Five refinements close precision gaps: term vectors alongside prose vectors, one-hop subclass expansion for high-confidence classes, full-text predicate search, density-ranked paragraph windows, and subclass-aware predicate matching. On the deduplication side, the pivotal design is the hard-conflict guard: two entities carrying

Load-bearing premise

The headline recall and zero-false-merge numbers are measured on two synthetic development documents with all thresholds tuned on them, so the load-bearing assumption is that behavior transfers to real production document streams.

What would settle it

Take a held-out set of real documents from the same domain with hand-built ground truth, run the pipeline with all thresholds frozen, and measure recall and count false merges; if recall falls well below 95 percent or any false merge appears, the zero-false-merge guarantee and the transferability premise fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A small, locally hosted 9B model can achieve ontology-aligned extraction and high search recall when prompted with a dynamically retrieved ontology slice, without relying on cloud-scale models.
  • Reclaiming roughly 10,500 tokens per extraction call from a ~11,200-token static catalog allows more source text to fit in a 32K-token window, reducing chunk fragmentation and cross-chunk merge errors.
  • The hard-conflict guard, combined with context-validated deduplication, makes the reported zero-false-merge result a property of the architecture rather than a lucky threshold choice.
  • Per-page OCR classification eliminates the information loss of binary text-or-OCR routing for mixed documents, a common real-world case.
  • Each pipeline stage addresses a distinct quality-defect class, so components can be adopted independently by other extraction systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The retrieval recipe (term vectors, subclass expansion, density-ranked windows, and subclass-aware predicate matching) is a general pattern for any embedding-based retrieval over a hierarchical schema, not only for intelligence documents.
  • The paper implicitly suggests that conservative synonym snapping (a 0.80 cosine floor) is the right trade-off to avoid semantic magnets; this is a testable hypothesis for other ontology-grounded extraction tasks.
  • The 58.8 percent OCR recall ceiling on the naval document suggests that the binding constraint for long-tail vocabulary is the vision model's reading ability, not the ontology layer; better OCR or multi-pass union would likely be the next gain.
  • If the architecture generalizes, it offers a template for compliance monitoring and investigative journalism: any governed-schema document stream becomes a queryable graph with auditable merge decisions.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper presents a production extraction layer for ontology-aligned knowledge graph construction from heterogeneous documents. The system consumes Kafka metadata, routes PDF/spreadsheet/Office/image content through format-specific handlers, performs two-phase LLM extraction with a locally hosted 4-bit Qwen3.5-9B model, and applies a five-stage refinement pipeline including deterministic cleaning, cross-chunk merging, a relationship second pass, six zero-inference deduplication algorithms, and an embedding-based resolution stage with a hard-conflict guard. The distinguishing component is ontology-guided extraction: a Neo4j ontology slice is retrieved live by embedding similarity and injected into the extraction prompt, reducing catalog overhead by roughly 94% relative to static domain slices. The paper reports that this design improves search recall from about 70% to 95% with no false merges, documents seven classes of upstream quality defects and their fixes, and includes an OCR stress test on a synthetic naval document plus a small public benchmark campaign comparing the local model to a cloud model.

Significance. If the reported gains hold beyond the development corpus, the paper makes a useful engineering contribution: live ontology retrieval for extraction-time grounding, a layered rule-based plus embedding deduplication pipeline with a hard-conflict guard, per-page OCR classification, and a detailed threshold reference (Appendix A) that goes beyond most systems papers. The honest documentation of seven real pipeline bugs and their fixes is a strength, as is the explicit listing of hand-tuned thresholds. However, the headline quantitative claims are not backed by independent validation: the evaluations use single synthetic development documents, thresholds are tuned on those same documents, no significance testing or confidence intervals appear, and the code/data are proprietary. The paper itself (Section 9, Section 12) acknowledges these limitations, but the abstract and conclusion present the numbers without those caveats, which is a serious mismatch between evidence and claim.

major comments (4)
  1. [§8.1, Table 17; §9; §12] The headline claim of search recall improving from roughly 70% to 95% rests on a single 45-page synthetic document (IR-001.pdf). Section 9 concedes that all thresholds (0.72 retrieval floor, 0.80 expansion trigger, similarity-scorer weights, indicator lists) were tuned on the development corpus, and Section 12 states the corpora are synthetic and cannot be redistributed. Table 17 does not describe the query set, the ground-truth duplicate pairs, or the exact protocol used to measure recall. As it stands, the reported improvement is a retrospective fit to the development document, not a predictive result. The authors should either provide a held-out evaluation with multiple documents and significance testing, or sharply qualify the claim in the abstract and Section 1 to 'on a development document.'
  2. [§6.2.4, §7.3.7, §8.1 Table 17] The claim of 'zero false merges' is not supported by the evidence. Section 7.3.7 documents 79 false-positive merge candidates produced by the abbreviation check, fixed only after evaluation; Table 16 shows 79 false-positive dedup candidates before the fix, while Table 17 reports false positives as 0/0. The corpus, being synthetic and lacking independent hard-negative labels, cannot substantiate a zero-false-merge guarantee. The claim should be reworded to 'no false merges observed in the two synthetic test documents after applying the fixes,' not presented as a general property of the system.
  3. [§4, Table 2; §4.1 Tables 4–5] The 94% catalog-overhead reduction is measured on a single 'representative document' (Table 2), with no distribution, sample size, or variance. Similarly, the retrieval refinements in Tables 4 and 5 are illustrated on one document/passage each. Without a corpus-level evaluation, the 94% figure is anecdotal. Please provide either a multi-document measurement with error bars or change the abstract and Section 1 to 'about 94% on a representative document.'
  4. [§8.2, §8.3, §9] The OCR ablation (10-page JFS document, chunk sizes 5 and 2) is a single run on one synthetic document, and the benchmark campaign (Section 8.3) uses samples of n=8–50 with the authors' own caveat that results are 'directional rather than statistically significant.' Despite this, the conclusion states that the pipeline 'tripled relationship coverage' and 'cut hallucinated entities from 174 to zero.' These are not robust claims in their current form. The authors should present the OCR and benchmark results as illustrative case studies and remove the generalizing language from the conclusion, or carry out the multi-pass and multi-document evaluation outlined in Table 20.
minor comments (3)
  1. [§8.1, Table 17] The row 'Graph false merges prevented: 0 → 4–8 per document' is ambiguous: before deduplication, zero false merges were 'prevented' because no deduplication existed; after deduplication, the claim is that 4–8 were prevented. Clarify the direction of the metric and reconcile with the 79 false-positive candidates reported in Table 16.
  2. [§3.2] The text says 'four domain-informed checks' but the list has four items; this is fine, but check 4's threshold of 'more than 30%' and 'fewer than 6 distinct types' should appear in Appendix A's threshold table for completeness.
  3. [§4.1] The phrase 'five systematic precision gaps' appears in the text, but the subsequent list and table contain G1–G5; this is internally consistent, but the introductory sentence says 'five' while Table 3 has five rows — good. However, Section 4.1 introduces 'five' and then Section 4 mentions 'four retrieval refinements' in the conclusion; please make the count consistent ('four' in the conclusion should be 'five' if G1–G5 are counted).

Circularity Check

2 steps flagged

Headline recall and zero-false-merge claims are retrospective fits on the development corpus; public benchmarks provide only partial independent support.

specific steps
  1. fitted input called prediction [Section 1.1; Section 8.1 (Table 17); Section 9; Section 12]
    "‘This methodology improved search recall from approximately 70% to 95% while maintaining a zero false-positive rate.’ ‘the similarity-scorer weights, retrieval thresholds (0.72 floor, 0.80 expansion trigger), and indicator-word lists are heuristics tuned on the development corpus; they have not been learned from labelled data and may not transfer across domains without re-tuning.’ ‘The evaluation corpora are synthetic intelligence-style documents created for system development.’"

    The headline 70→95% recall and zero-false-merge claims are measured on the same two synthetic documents on which every threshold (0.72 retrieval floor, 0.80 expansion trigger, similarity weights, indicator lists) was tuned. The evaluation set is the tuning set, so the improvement is a retrospective fit rather than an out-of-sample result. Section 9 concedes no significance testing and that the values may not transfer; Section 12 says the corpora are synthetic and non-redistributable, removing an external check. The strongest empirical claim therefore reduces to a measure of its own fitting inputs.

  2. fitted input called prediction [Section 4.1]
    "‘operational evaluation on intelligence-domain documents exposed five systematic precision gaps ... Each gap traces to one design decision in the baseline retriever, and each admits a targeted refinement ... Measured Effect. The refinements change which classes reach the prompt, and the change is visible immediately on real documents.’"

    The G1–G5 retrieval refinements were designed after observing failures on the evaluation documents, and their reported effect (e.g., Table 5: baseline ‘3 predicates’ vs refined ‘68’) is measured on those same documents. This is iterative tuning on the test set presented as a measured improvement; the gain is partly engineered into the metric rather than independently predicted. It is secondary to the headline recall claim but shows the same tuning/evaluation overlap.

full rationale

Most of the paper is a system description with engineering heuristics; there is no formal derivation whose conclusion is assumed as an input. The standard methods (Jaro-Winkler, Double Metaphone, FAISS, RAG, entity-resolution blocking) are cited externally, not self-cited, and no uniqueness/ansatz is smuggled in via the authors' prior work. The load-bearing circularity is empirical: the headline recall improvement and zero-false-merge claim are measured on two synthetic development documents after all retrieval/merging thresholds were tuned on that same corpus. The paper states this plainly in Section 9 and Appendix A, and it also says the corpora are synthetic and cannot be redistributed (Section 12). So the 70→95% recall number is a retrospective fit, not a prediction with independent support. The public benchmark campaign (Section 8.3) is less circular and provides some independent calibration, but it does not support the headline recall/overhead/zero-false-merge claims. Hence score 6: one or more central 'results' reduce to fits on their own inputs, but the architecture retains independent content and the paper is transparent about the limitation.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 0 invented entities

Every threshold in the pipeline is hand-set and tuned on the same development corpus used for evaluation (Section 9, Appendix A); the paper itself warns these 'have not been learned from labelled data and may not transfer across domains without re-tuning.' The quantitative claims therefore rest on a large set of free parameters, not on a parameter-free derivation or independent benchmark.

free parameters (9)
  • PageRank penalty λ = 0.4
    Down-weights generic classes in ontology retrieval (Section 4).
  • Retrieval cosine floor = 0.72
    Minimum adjusted score for a class/predicate to be injected; empirically tuned (Section 4, Appendix A).
  • Subclass expansion trigger/score = 0.80 / 0.78
    Classes scoring ≥0.80 trigger one-hop subclass expansion; children injected at 0.78 (Section 4.1).
  • Predicate embedding snap floor = 0.80
    Only predicate synonyms ≥0.80 cosine snap to ontology vocabulary (Section 4.2).
  • Source-text variant miner threshold τ = 0.85
    Ratcliff/Obershelp similarity threshold for accepting spelling variants (Section 6.2.2).
  • Context-aware dedup thresholds = 0.75 (candidate), 0.95 (auto-merge), 0.90–0.95 (review)
    Decision surface in Section 6.2.4.
  • Embedding resolution thresholds = 0.85 auto-merge, 0.60–0.85 review, 0.50 codename
    Decision engine thresholds (Section 6.5).
  • Similarity composite weights = w=(0.30,0.25,0.20,0.15,0.10)
    Five-signal weighted composite for name similarity (Section 6.2.3).
  • OCR page classifier thresholds = image coverage >0.15, chars >50, drawing >200, route ratios 0.80/0.20
    Per-page classification cascade and PDF routing (Section 5.1).
axioms (4)
  • domain assumption A curated ontology graph with class/predicate definitions and precomputed embeddings is available and maintained.
    The entire ontology-guided retrieval stage (Section 4) depends on this graph; the paper does not describe its construction or quality.
  • domain assumption The locally hosted 9B model (Qwen3.5-9B-AWQ-4bit) is capable of following ontology-guided extraction prompts well enough for production; the benchmark shows it trails Gemini by ~6 points on entity spans.
    The system's extraction quality depends on model competence; local-vs-cloud gap in Section 8.3 is directional only.
  • domain assumption The hard-conflict guard on discriminator attributes (DOB, nationality, ID) is a sufficient anti-merge safeguard; i.e., extracted discriminator values are accurate.
    Section 6.5 asserts the guard prevents false merges, but the paper does not evaluate its accuracy or the reliability of the discriminator fields.
  • standard math Standard similarity and retrieval primitives (Jaro–Winkler, Double Metaphone, Ratcliff/Obershelp, PageRank, FAISS) behave as documented.
    Used in Sections 4 and 6; these are standard and need no new justification.

pith-pipeline@v1.3.0-alltime-deepseek · 33746 in / 13151 out tokens · 113300 ms · 2026-08-03T00:44:08.597627+00:00 · methodology

0 comments
read the original abstract

Large language models extract entities and relationships from unstructured documents fluently but inconsistently: type vocabularies fracture across documents, the same person surfaces under several name variants, relationships duplicate, and distinct individuals who share a name risk silent conflation. This paper presents the design, implementation, and empirical refinement of a production extraction layer that converts a live document stream into a validated knowledge graph aligned to a formal ontology. The system consumes document metadata from Kafka, routes PDF, spreadsheet, Office, and image content through handlers built for each format, and extracts entities and relationships in two passes using a locally hosted Qwen3.5-9B model tuned on the ontology. Its distinguishing component is ontology-guided extraction: the relevant slice of a curated ontology is retrieved live from a graph database by embedding similarity and injected into the extraction prompt, reducing catalog overhead by about 94 percent relative to static domain slices. Extracted results then pass through a refinement pipeline of five stages: deterministic cleaning, merging across chunks, a second pass for relationships, six deduplication algorithms that require no model inference, and an embedding resolution subsystem whose conflict guard no similarity score can override. Evaluation on intelligence corpora improved search recall from roughly 70 to 95 percent with no false merges, and corrected seven classes of silent quality defect, ranging from a bug that truncated source text by a single character to the systematic duplication of entities that carried title prefixes.

Figures

Figures reproduced from arXiv: 2607.28662 by Kevin Lewis, Kundeshwar Pundalik, Vaibhav Dangaich.

Figure 1
Figure 1. Figure 1: High-level architecture of the extraction layer. The Kafka consumer routes each document by MIME [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ontology-guided extraction via live graph retrieval, read as a U-shaped flow: the left column retrieves [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The refined retrieval pipeline. The five refinements (red badges) slot into the baseline of Fig. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Multi-format routing. The consumer dispatches each record by MIME type and extension to a [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Per-page classification on a mixed document. Each page carries the signal that decided it: born-digital [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Schema evolution. The legacy schema (left) carried one alias per entity and a bare disambiguation [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Algorithmic alias expansion. A single extracted name deterministically radiates to its abbreviative [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Source-text variant mining in action. The miner scans the raw document for word pairs matching the [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Five-signal composite scoring for “John Doe” vs. “Jon Doe”. The Double Metaphone and [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The context-validated decision surface. Name similarity alone never merges: only the top-left cell [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Qualifier-preserving relationship deduplication. Three extractions of the same edge (bare, dated, and [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Sequence-ordered output under parallel processing. Three workers finish out of order: the short [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Phase 3 orchestration as a guarded composition [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Embedding-based entity resolution. Entities are embedded and blocked by the union of a FAISS [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: High-level overview of the five-stage extraction pipeline. A document enters at the left and flows [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Stage 1: LLM Extraction. The document is split into overlapping chunks; two sequential LLM calls [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Stage 2: Per-Chunk Finalization. Three deterministic cleaning passes normalise relationship types, [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Stage 3: Cross-Chunk Merging. Title-aware normalisation produces canonical keys; entities sharing [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Stage 4: Relationship Second Pass. The merged entity catalog is batched and sent to the LLM for [PITH_FULL_IMAGE:figures/full_fig_p028_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Stage 5: Post-Extraction Enhancements. Four algorithms enrich the knowledge graph with expanded [PITH_FULL_IMAGE:figures/full_fig_p029_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Pipeline output: a validated JSON knowledge graph containing typed entities, evidence-linked [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: A single output record rendered as the graph fragment it becomes. The entity carries six aliases [PITH_FULL_IMAGE:figures/full_fig_p030_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Summary of the principal quantitative results from the deduplication and ontology-grounding [PITH_FULL_IMAGE:figures/full_fig_p035_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Four headline dimensions on the JFS document. The legacy pipeline misclassifies all naval vessels as [PITH_FULL_IMAGE:figures/full_fig_p036_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: Ground-truth vessel-class recall vs. chunk size (pages per LLM call) on the 10-page JFS document. [PITH_FULL_IMAGE:figures/full_fig_p037_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: OCR token-F1 across three datasets and three engines. The local model matches the cloud model on [PITH_FULL_IMAGE:figures/full_fig_p038_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: Ontology results (local model). High conformance and near-zero hallucination coexist with low [PITH_FULL_IMAGE:figures/full_fig_p039_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: End-to-end accuracy on a ground-truth intelligence document across six dimensions (overall [PITH_FULL_IMAGE:figures/full_fig_p039_28.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 2 linked inside Pith

  1. [1]

    Kreps, J., Narkhede, N., & Rao, J. (2011). Kafka: A distributed messaging system for log processing. InProceedings of the NetDB(Vol. 11, pp. 1–7)

    2011

  2. [2]

    InProceedingsofthe29thSymposiumonOperatingSystemsPrinciples(SOSP)(pp.611–626)

    Kwon,W.,Li,Z.,Zhuang,S.,Sheng,Y.,Zheng,L.,Yu,C.H.,Gonzalez,J.,Zhang,H.,&Stoica, I.(2023).EfficientmemorymanagementforlargelanguagemodelservingwithPagedAttention. InProceedingsofthe29thSymposiumonOperatingSystemsPrinciples(SOSP)(pp.611–626)

    2023

  3. [3]

    Pan,S.,Luo,L.,Wang,Y.,Chen,C.,Wang,J.,&Wu,X.(2024).Unifyinglargelanguagemodels and knowledge graphs: A roadmap.IEEE Transactions on Knowledge and Data Engineering, 36(7), 3580–3599

    2024

  4. [4]

    Zhu, Y., Wang, X., Chen, J., Qiao, S., Ou, Y., Yao, Y., Deng, S., Chen, H., & Zhang, N. (2024). LLMs for knowledge graph construction and reasoning: Recent capabilities and future opportunities.World Wide Web,27(5), 58

    2024

  5. [5]

    (2024).Pydantic(Version 2) [Computer software].https://github.com/ pydantic/pydantic

    Colvin, S. (2024).Pydantic(Version 2) [Computer software].https://github.com/ pydantic/pydantic

    2024

  6. [6]

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive NLP tasks. InAdvances in Neural Information Processing Systems(Vol. 33, pp. 9459–9474)

    2020

  7. [7]

    Christophides, V., Efthymiou, V., Palpanas, T., Papadakis, G., & Stefanidis, K. (2020). An overview of end-to-end entity resolution for big data.ACM Computing Surveys,53(6), 1–42

    2020

  8. [8]

    Papadakis,G.,Skoutas,D.,Thanos,E.,&Palpanas,T.(2020).Blockingandfilteringtechniques for entity resolution: A survey.ACM Computing Surveys,53(2), 1–42

    2020

  9. [9]

    W., & Metzener, D

    Ratcliff, J. W., & Metzener, D. E. (1988). Pattern matching: The gestalt approach.Dr. Dobb’s Journal,13(7), 46–51

    1988

  10. [10]

    Knight, K., & Graehl, J. (1998). Machine transliteration.Computational Linguistics,24(4), 599–612

    1998

  11. [11]

    N., Kaiser, Ł., & Polosukhin,I.(2017).Attentionisallyouneed.InAdvancesinNeuralInformationProcessing Systems(Vol

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin,I.(2017).Attentionisallyouneed.InAdvancesinNeuralInformationProcessing Systems(Vol. 30). 41

    2017

  12. [12]

    B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020). Language models are few-shot learners. In Advances in Neural Information Processing Systems(Vol. 33, pp. 1877–1901)

    2020

  13. [13]

    Wei, X., Cui, X., Cheng, N., Wang, X., Zhang, X., Huang, S., Xie, P., Xu, J., Chen, Y., Zhang, M., Jiang, Y., & Han, W. (2023). ChatIE: Zero-shot information extraction via chatting with ChatGPT.arXiv.https://arxiv.org/abs/2302.10205

    Pith/arXiv arXiv 2023

  14. [14]

    Wang, S., Sun, X., Li, X., Ouyang, R., Wu, F., Zhang, T., Li, J., & Wang, G. (2025). GPT-NER:Namedentityrecognitionvialargelanguagemodels.InFindingsoftheAssociation for Computational Linguistics: NAACL 2025(pp. 4257–4275)

    2025

  15. [15]

    InFindings of the Association for Computational Linguistics: EMNLP 2021(pp

    HuguetCabot,P.-L.,&Navigli,R.(2021).REBEL:Relationextractionbyend-to-endlanguage generation. InFindings of the Association for Computational Linguistics: EMNLP 2021(pp. 2370–2381)

    2021

  16. [16]

    Wadhwa, S., Amir, S., & Wallace, B. (2023). Revisiting relation extraction in the era of large language models. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL)(pp. 15566–15589)

    2023

  17. [17]

    Gruber, T. R. (1993). A translation approach to portable ontology specifications.Knowledge Acquisition,5(2), 199–220

    1993

  18. [18]

    Stanford Digital Library Technologies Project

    Page,L.(1998).ThePageRankcitationranking: Bringingordertotheweb(TechnicalReport). Stanford Digital Library Technologies Project

    1998

  19. [19]

    P., & Sunter, A

    Fellegi, I. P., & Sunter, A. B. (1969). A theory for record linkage.Journal of the American Statistical Association,64(328), 1183–1210

    1969

  20. [20]

    K., Ipeirotis, P

    Elmagarmid, A. K., Ipeirotis, P. G., & Verykios, V. S. (2007). Duplicate record detection: A survey.IEEE Transactions on Knowledge and Data Engineering,19(1), 1–16

    2007

  21. [21]

    Mudgal, S., Li, H., Rekatsinas, T., Doan, A., Park, Y., Krishnan, G., Deep, R., Arcaute, E., & Raghavendra, V. (2018). Deep learning for entity matching: A design space exploration. In Proceedings of the 2018 International Conference on Management of Data (SIGMOD)(pp. 19–34)

    2018

  22. [22]

    Brunner, U., & Stockinger, K. (2020). Entity matching with transformer architectures: A step forward in data integration. InProceedings of the 23rd International Conference on Extending Database Technology (EDBT)(pp. 463–473)

    2020

  23. [23]

    InSixth IEEE International Conference on Data Mining Workshops (ICDM)(pp

    Christen,P.(2006).Acomparisonofpersonalnamematching: Techniquesandpracticalissues. InSixth IEEE International Conference on Data Mining Workshops (ICDM)(pp. 290–294)

    2006

  24. [24]

    Cohen, W., Ravikumar, P., & Fienberg, S. (2003). A comparison of string distance metrics for name-matching tasks. InProceedings of the IJCAI-2003 Workshop on Information Integration on the Web (IIWeb)(pp. 73–78)

    2003

  25. [25]

    Navarro, G. (2001). A guided tour to approximate string matching.ACM Computing Surveys, 33(1), 31–88

    2001

  26. [26]

    Winkler, W. E. (1990). String comparator metrics and enhanced decision rules in the Fellegi–Sunter model of record linkage. InProceedings of the Section on Survey Research Methods, American Statistical Association(pp. 354–359)

    1990

  27. [27]

    Philips, L. (2000). The double metaphone search algorithm.C/C++ Users Journal,18(6), 38–43

    2000

  28. [28]

    T., Condon, S

    Freeman, A. T., Condon, S. L., & Ackerman, C. M. (2006). Cross linguistic name matching in English and Arabic. InProceedings of the Human Language Technology Conference of the NAACL(pp. 471–478)

    2006

  29. [29]

    Shen,W.,Wang,J.,&Han,J.(2015).Entitylinkingwithaknowledgebase: Issues,techniques, and solutions.IEEE Transactions on Knowledge and Data Engineering,27(2), 443–460. 42

    2015

  30. [30]

    Cucerzan, S. (2007). Large-scale named entity disambiguation based on Wikipedia data. InProceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning (EMNLP-CoNLL)(pp. 708–716)

    2007

  31. [31]

    Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using Siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)(pp. 3982–3992)

    2019

  32. [32]

    Artifex Software.(2024).PyMuPDFdocumentation.https://pymupdf.readthedocs.io/

    2024

  33. [33]

    Johnson, J., Douze, M., & Jégou, H. (2019). Billion-scale similarity search with GPUs.IEEE Transactions on Big Data,7(3), 535–547

    2019

  34. [34]

    F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P

    Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts.Transactions of the Association for Computational Linguistics,12, 157–173

    2024

  35. [35]

    Yang,A.,Li,A.,Yang,B.,Zhang,B.,Hui,B.,Zheng,B.,...&Qiu,Z.(2025).Qwen3technical report.arXiv.https://arxiv.org/abs/2505.09388

    Pith/arXiv arXiv 2025

  36. [36]

    vaguely related

    Wang, G., Koshy, J., Subramanian, S., Paramasivam, K., Zadeh, M., Narkhede, N., Rao, J., Kreps, J., & Stein, J. (2015). Building a replicated logging system with Apache Kafka. Proceedings of the VLDB Endowment,8(12), 1654–1655. A Threshold reference Table 22 consolidates every tunable threshold in the extraction and resolution pipeline, the subsystem it g...

    2015