PaperlayerView exact revision on arXiv

Source-linked AI summary

ArticleMiner: Ontology-Guided Knowledge Graph Construction from Scientific Publications

Md Abrar Jahin, Craig A. Knoblock, Jay Pujara

arXiv:2609.25607v1cs.AI

TL;DR

Scientific observations in publications need contextual interpretation beyond recovering table layout, while many table-interpretation methods assume usable tables and publication-level extraction systems are often domain-specific. ArticleMiner combines shared publication-level evidence processing with bounded task-specific ontology modules, then evaluates four task instantiations. Its reported raw-publication results favor the complete pipeline over same-LLM few-shot prompting, although the comparisons do not isolate ontology grounding.

  • Problem

    Interpreting quantitative observations requires context such as headers, units, and captions, while semantic table systems generally assume usable tables and publication-level extraction systems are typically domain-specific.

    Source: §1 · p. 1

  • Method

    ArticleMiner combines shared publication-level orchestration with bounded human-authored ontology modules that specify canonical terms, mappings, rules, validation, identity, and RDF bindings.

    Source: §1 · p. 2, §2.1 · p. 3

  • Results

    Across four authored tasks, reported raw-publication results favor the complete pipeline over same-LLM few-shot prompting, with uncertainty on the two smaller benchmarks.

    Source: §4.1 · p. 10, §7 · p. 13

  • Takeaways & Limitations

    The framework provides a reusable interface for acquiring, interpreting, and reconciling evidence under bounded domain constraints.

    Source: §7 · p. 13

    Full-system comparisons change multiple stages, and GeoChem also differs in supplementary-file access and identifier handling, so they do not isolate ontology grounding.

    Source: §6 · p. 13

Abstract

from arXiv · show

Scientific papers keep much of their quantitative content in tables and supplementary files, where a number means something only through its header, caption, unit, analytical method, and the conventions of its field. Recovering the rows and columns of a table is therefore not the same as recovering the scientific fact it reports. Most semantic table-interpretation methods assume that a clean table is already available and subsequently map its cells or columns to ontology terms, whereas most publication-level extraction systems are designed for a single domain. We study a middle path: a shared process that reads a paper and its supplementary files, gathers evidence from several parsers and a language model, and reconciles that evidence, while a bounded human-authored task module for each task supplies the domain meaning. The module lists the canonical names the graph may use, the surface forms that map to them, a small set of derivation rules and validity constraints, an identity key, and the bindings used to write RDF. It defines what a task is allowed to emit; it does not try to list every convention of a field. We build four such modules (for drug-discovery chemistry, materials science, machine learning, and mineral geochemistry) in the ArticleMiner framework, and evaluate them on 163 papers, including a new geochemistry benchmark with expert-curated ground truth. In comparisons against a same-LLM few-shot baseline, the point estimates favor ArticleMiner on all four tasks, with uncertainty on the two smaller benchmarks. The geochemistry comparison also includes access to supplementary files, so its improvement cannot be attributed to domain guidance alone.

1. Introduction

Scientific observations require contextual interpretation, not just table-layout recovery, to preserve their meaning in structured records. ArticleMiner combines shared publication-level processing with bounded, human-authored task modules and evaluates this approach across four domains.

1. Introduction

  • Interpreting a table observation requires its header, caption, unit, and other context to preserve distinctions such as censored values versus measured values.For example, <0.5 under an Au (ppm) header denotes a below-detection observation, not a measured value.Source: §1 · p. 1
  • Semantic table-to-KG systems generally assume a usable table representation, while publication-level extraction systems are typically engineered for one domain or target relation [1–7].ArticleMiner investigates an explicit task-module interface for reusing publication-level evidence acquisition and extraction across domains.Source: §1 · p. 1
  • The study asks how domain guidance affects extraction quality, how evidence sources can be reconciled with provenance, and how acquisition failures and interpretation errors contribute to remaining errors.Same-LLM comparisons and component ablations address the first question; reconciliation and backend ablations address the second; PDF/pre-parsed comparisons and error analysis address the third.Source: §1 · pp. 1–2
  • ArticleMiner pairs shared publication-level orchestration with bounded ontology modules and evaluates four task instantiations across 163 papers.The modules define permissible outputs without attempting to axiomatize the entire field.Source: §1 · p. 2
  • The contributions are ontology-constrained extraction, a module and evidence interface for provenance-preserving reconciliation, and an evaluation distinguishing full-system comparisons from component ablations.The evaluation covers four task instantiations over 163 papers, including raw-PDF and pre-parsed settings and matched same-LLM baselines.Source: §1 · p. 2

2. Problem Formulation and Method

ArticleMiner pairs shared publication-level orchestration with a bounded task module that defines permissible outputs. Its method uses contextual evidence and declared rules to preserve scientific meaning through extraction, validation, reconciliation, and serialization.

2. Problem Formulation and Method

  • A bounded task module defines permissible outputs within a shared pipeline that parses, extracts, self-corrects, reconciles, and serializes publication evidence.The modules provide task semantics without attempting to axiomatize the entire field.Source: §2.1 · p. 2, §2 · p. 2, §2.3 · p. 5

2.2. Geochemistry Instantiation

  • GeoChem maps curated mineral and method aliases to canonical labels and derives deposit parents by lookup in the adopted USGS CMMI classification.An unlisted surface form is retained for validation rather than assigned an unsupported canonical label.Source: §2.2 · p. 4
  • GeoChem distinguishes detection limits, below-detection values without stated limits, and unmeasured values; ambiguous units are not inferred from magnitude alone.An explicitly labeled wt% value can be converted to ppm, while validation checks terms, units, required fields, and numerical plausibility without guaranteeing correct sample attachment.Source: §2.2 · pp. 4–5

2.3. Shared Orchestration

  • Five PDF backends contribute evidence because individual parsers miss or corrupt different table layouts; supplementary workbook and sheet metadata can also inform extraction.MinerU lacked OCR weights in the evaluation environment, so its results do not characterize a fully enabled installation.Source: §2.3 · p. 5
  • Extraction uses publication context to connect cells with their sample, element, unit, method, and measurement convention before normalization.For example, interpreting Au<0.5 requires associating it with ppm, sample BR-12, and ICP-MS; domain guidance narrows labels but does not replace source evidence.Source: §2 · p. 2, §2.3 · p. 5
  • Self-correction uses expected sample counts and targeted parser retries to address missing samples or structural failures, but these diagnostics do not establish semantic correctness.The geochemistry retry lowers the recognized-element-column threshold and replaces an earlier set only when it recovers more samples.Source: §2.3 · p. 6
  • Reconciliation derives, canonicalizes, and validates candidates before grouping them by task identity; backend support measures identity recovery, not factual correctness.The JSON-LD export separates sample identity, measurements, and source metadata, but export and downstream RDF validation are separate from the evaluated tuple-level task.Source: §2.3 · pp. 6–7

2.4. A Worked Example

  • Across geochemistry, bioactivity, glass, and machine-learning examples, preserving contextual bounds, units, and identifiers matters because schema-valid records can still express the wrong scientific claim.The released JSON-LD structure keeps sample identity, measurements, and source metadata distinct.Source: §2.4 · p. 7, §2.3 · p. 7

3. Evaluation Design

The evaluation spans four benchmarks and distinguishes task-specific scoring from controlled same-LLM comparisons. It treats the two smallest datasets as case evidence and labels secondary analyses accordingly.

3. Evaluation Design

  • The four benchmarks vary in input modality, ontology size, tuple granularity, and reliance on supplementary evidence; the two smallest are treated as case evidence, not population-level confirmation.The 163 papers are not 163 independent draws of domain portability.Source: §3 · p. 7

3.1. Benchmark Datasets

  • GeoChem adds 28 expert-curated mineral-geochemistry papers with 1,236 sample identifiers and 87,759 non-null elemental measurements; three external benchmarks cover chemistry, glass composition, and machine-learning tables.GeoChem records are matched to ground truth using sample identifiers.Source: §3.1 · pp. 7–8

3.2. Four-Tier Evaluation for GeoChem

  • GeoChem’s composite S weights numerical accuracy (T2) at 0.40, metadata (T1) at 0.30, and sample recovery (T3) and missing-value agreement (T4) at 0.15 each.T2 and T4 are conditional on row matching, so they do not independently measure extraction recall or full three-way null semantics.Source: §3.2 · p. 8

3.3. Evaluation Metrics

  • External benchmarks use precision, recall, and strict tuple-F1, whereas GeoChem’s primary four-tier score is distinct; sample-row P/R/F1 is reported for cross-benchmark comparability.The reported attribute-level token-F1 results of Bai et al. are context only and not directly comparable to strict tuple-F1.Source: §3.3 · pp. 8–9

3.4. Baselines, Configurations, and Statistical Protocol

  • B1 uses the same backbone on raw-PDF text with fixed three-exemplar prompting, but GeoChem’s spreadsheet-aware comparison also differs in evidence access, schema presentation, and identifier handling.Table 4(b) reports mean per-paper scores with percentile bootstrap intervals from 1,000 resamples, paired by publication identifier.Source: §3.4 · p. 9

4. Results

Across raw-publication comparisons, ArticleMiner’s point estimates exceeded same-LLM few-shot results on four tasks and five closed backbones, while component and backend effects varied by task. Error analysis distinguishes over-extraction, omission, attribute mismatch, and near miss.

4.1. Same-Backbone Comparison and Uncertainty

  • 8.8 points on ChemTables, 14.4 on MLTables, 15.5 on DiSCoMaT, and 62.3 sample-F1 points on GeoChem were the reported mean gains over same-LLM few-shot.The smaller ChemTables and MLTables samples have substantial uncertainty; GeoChem includes supplementary spreadsheets, so its improvement is not attributable to ontology guidance alone.Source: §4.1 · pp. 9–10, Table 4 · p. 10
  • 81.0 tuple-F1 on raw DiSCoMaT PDFs versus 94.7 on pre-parsed tables was reported for Opus 4.6; pre-parsed input scored higher across three external benchmarks.This gap indicates acquisition losses but does not isolate parsing from changes in evidence representation or prompting.Source: §4.1 · p. 9

4.2. Contribution of the Domain Specification: Component Ablations

  • Removing ontology grounding reduced tuple-F1 by 16.3 points on DiSCoMaT, 7.7 on MLTables, and 2.2 on ChemTables; component effects also reversed across backbones.Under Sonnet, removing vision reduced ChemTables from 32.1 to 14.6 and MLTables from 52.2 to 4.8; these diagnostics do not establish a universally optimal component set.Source: §4.2 · p. 10

4.3. Acquisition, Backends, and Backbone Sensitivity

  • 71.7 F1 for Docling alone on DiSCoMaT contrasted with 43.5 without Docling and 71.0 for all-five fusion; on GeoChem, all-five fusion reached 54.3 F1.On ChemTables and MLTables, Docling-only scores exceeded all-five fusion (26.2 vs. 20.8; 56.3 vs. 47.6), making backend choice task-dependent.Source: §4.3 · pp. 10–11, Table 5 · p. 11
  • 52–58% cross-LLM agreement on external benchmarks contrasted with 75% on GeoChem; the ChemTables exemplar-count analysis plateaued near three examples.The GeoChem open-weight comparison used a different 26-paper subset and a 512-token stage cap, so it does not isolate model capacity.Source: §4.3 · p. 11

4.4. Error Analysis

  • Errors are classified as over-extraction, omission, attribute mismatch, or near miss, and model choice mostly shifts the precision–recall balance.Sonnet over-extracts on ChemTables and MLTables but under-extracts on DiSCoMaT; GPT-4o under-extracts everywhere, while validation cannot recover candidates never proposed.Source: §4.4 · pp. 11–12
  • Omissions made up 92.1% of Sonnet’s 1,721 DiSCoMaT error events, versus 75.5% of the few-shot baseline’s 2,697; the larger share did not mean more omissions in absolute terms.Over-extraction instead accounted for 57.0% and 59.9% of ArticleMiner’s ChemTables and MLTables errors, respectively.Source: §4.4 · p. 12

5. Related Work

Related work covers semantic annotation of recovered tables, domain-specific scientific extraction, and downstream RDF construction. ArticleMiner combines publication-level evidence acquisition with an explicit domain module before a structured source exists.

Table-to-KG and semantic table annotation

  • SemTab, DAGOBAH, and SAND annotate cells, columns, or properties in already-recovered tables, whereas ArticleMiner couples parsing with interpretation of surrounding evidence [1–3, 9, 10].The cited systems include domain-independent interpretation and DBpedia table matching.Source: §5 · p. 12

Scientific IE and domain-specific KGs

  • ChemDataExtractor, DiSCoMaT, and MinMod demonstrate domain knowledge in scientific extraction but are engineered around one domain, source, or target schema.ArticleMiner externalizes domain knowledge into a module over shared orchestration, including executable rules and, when needed, a classifier.Source: §5 · p. 12

LLM-based table extraction and document parsing; Semantic lifting and ontology-guided KG construction

  • Unlike systems primarily operating over structured or recovered tables [4, 11–15] and RDF lifting from structured sources, ArticleMiner recovers units, null values, identifiers, and context before a structured source exists.The passage also notes that no single layout-aware parser is reliable across scientific layouts [16] [17] [18].Source: §5 · p. 12

6. Discussion and Threats to Validity

Across the evaluated tasks, domain guidance contributes to extraction quality, but gains depend on task and backbone. Acquisition failures remain an important limitation, and the evidence supports task-specific rather than universal policies.

6. Discussion and Threats to Validity

  • Domain guidance contributes to extraction quality, but gains vary by task and backbone; adding extractors does not consistently improve accuracy.Acquisition gaps and over-extraction also remain sources of error.Source: §6 · p. 12
  • A new task requires vocabulary and mappings, derivation and validation code, prompts, a row schema, identifiers, and RDF bindings; geochemistry also needs a deposit classifier.The four modules demonstrate orchestration reuse, not zero-shot transfer or a measured reduction in authoring time.Source: §6 · p. 12
  • Define the target observation before expanding vocabulary; sample-name-only identity keys can merge analytical methods, while overly specific keys can retain duplicates.The effect of this recommendation on authoring effort remains unevaluated.Source: §6 · p. 13
  • GeoChem’s composite score can rise without better end-to-end recall because matching and missing-metadata rules may award credit; no direct downstream KG query-quality measure is reported.The weights can change the ranking of nearby variants, so per-tier scores are needed.Source: §6 · p. 13
  • Full-system comparisons change multiple stages, and GeoChem also differs in supplementary-file access and identifier handling, limiting causal interpretation of ontology grounding.Post-hoc configurations, ground-truth snapshots, token limits, and unavailable MinerU OCR weights further restrict interpretation.Source: §6 · p. 13
  • The 163 papers cover four manually authored tasks, with small ChemTables and MLTables samples; GeoChem relies heavily on machine-readable supplements and records are not resolved across papers.Small samples yield imprecise estimates; the paired summaries do not establish formal significance or absence of an effect, and repeated-run variability was insufficiently characterized.Source: §6 · p. 13

7. Conclusion

ArticleMiner offers a reusable interface for acquiring, interpreting, and reconciling publication evidence under bounded task-specific constraints. Raw-publication results favor the complete pipeline over same-LLM few-shot prompting across four tasks, while general adaptation cost and downstream graph quality remain open evaluation needs.

7. Conclusion

  • Across four authored tasks, reported raw-publication results favor the complete pipeline over same-LLM few-shot prompting, with substantial task dependence in component and parser analyses.The framework contributes a reusable interface for acquiring, interpreting, and reconciling evidence under bounded domain constraints.Source: §7 · p. 13
  • General adaptation cost, controlled supplementary-aware baselines, and downstream graph quality remain open evaluation needs.Future work includes cross-paper entity resolution, adaptive evidence acquisition, and expert-reviewed module expansion from logged failures.Source: §7 · p. 13
  • The release includes source code, ontology modules, prompts, schemas, outputs, ablation grids, and expert-contributed GeoChem ground truth under CC-BY-4.0.Few-shot outputs cover Sonnet 4.6 and open backbones; other closed-backbone results require regeneration with the released runner.Source: §7 · p. 13

Declaration of use of Generative AI

The paper discloses generative AI use for both research instrumentation and manuscript assistance. It then distinguishes ontology-module roles, illustrates cross-domain contextual interpretation, and specifies benchmark output schemas.

Declaration of use of Generative AI

  • Generative AI served as both research instrumentation in the pipeline and manuscript assistance for language editing; authors reviewed and edited assisted text.The listed models include Claude, GPT-4o, Gemini, and open 7–8B models.Source: Declaration of use of Generative AI · p. 14

A. The Three Integration Points of the Ontology Module

  • The ontology module guides proposed fields at extraction, canonicalizes and validates records at reconciliation, and fixes graph representation at serialization.These roles are functionally distinct and are not separate empirical ablations.Source: A. The Three Integration Points of the… · p. 17

B. Cross-Domain Worked Example Analogs

  • Cross-domain examples show context-dependent meanings: mol% differs from wt%, IC50 >10 µM is a lower bound, and resolving “Ours” needs source context.The examples distinguish model-led contextual interpretation from deterministic module operations; they are schematic, not additional evaluated outputs.Source: B. Cross-Domain Worked Example Analogs · pp. 17–18

C. Output Schemas

  • The benchmarks define distinct tuple schemas, while GeoChem’s 209-column gold schema comprises 63 metadata/provenance fields and 146 element columns.Evaluation uses the named-field projection, so additional emitted fields are ignored rather than penalized.Source: C. Output Schemas · p. 18

C.1. GeoChem scoring details

GeoChem scoring combines field-level string and numerical measures with sample matching and null agreement. Its alignment fallback uses numerical overlap, so the resulting sample-F1 is not equivalent to strict tuple matching.

C.1. GeoChem scoring details

  • T1 scores strings by the maximum of normalized token-Jaccard and sequence similarity; T2 scores numerical fields by relative error, with full credit through 5% error and zero at 100% or more.A zero gold value receives full credit only within absolute error 0.001.Source: §C.1 · p. 18
  • For an ordinary gold value, T2 scores a predicted sentinel 0 and another predicted value 0.3; it averages scored fields within each matched row, then averages matched rows.T4 similarly averages null agreement over gold-null fields.Source: §C.1 · p. 18
  • When fewer than 30% of gold rows match by name, fallback alignment uses row order and numerical overlap; the resulting T3 sample-F1 is not strict tuple matching.Compatible rows require at least 30% of shared positive elemental values to agree within 10%.Source: §C.1 · p. 18

D. Per-Domain External-Benchmark Detail

Per-domain results show distinct strengths and constraints across the four benchmarks, alongside task-specific vocabularies and extraction conventions. The appendix also documents prompt choices and backbone-sensitive effects on PDF-only tasks.

D. Per-Domain External-Benchmark Detail

  • 81.0 and 79.3 tuple-F1 on DiSCoMaT were reported for Opus 4.6 and Sonnet 4.6, versus same-backbone few-shot scores of 59.5 and 64.5.Sonnet had 98.2% precision and 66.4% recall; 32 of 111 papers had no predictions and remained in the denominator.Source: D. Per-Domain External-Benchmark Detail · p. 18
  • 55.9 tuple-F1 for Opus 4.6 on MLTables exceeded its few-shot baseline’s 41.6, while the 15-paper sample limits these observations.Separately, Sonnet’s paired means were 52.2 versus 37.8, a reported paired difference of 14.4.Source: D. Per-Domain External-Benchmark Detail · p. 18
  • 32.1 tuple-F1 for the full Sonnet ChemTables configuration compared with 15.8 for few-shot prompting; disabling paper intelligence post hoc raised the score to 44.6.The 9-paper paired summary has wide intervals and an incompletely documented configuration, so it is not a confirmed paired test of the full pipeline.Source: D. Per-Domain External-Benchmark Detail · pp. 18–19
  • GeoChem’s closed-backbone composite scores ranged from 75.5 to 76.3 on 28 entries; its separate 25-entry paired analysis reported a Sonnet mean sample-F1 of 63.2.The estimands and denominators differ, and supplementary-file access and sample-identifier handling contribute to the gap from PDF-only few-shot prompting.Source: D. Per-Domain External-Benchmark Detail · p. 19

E. Ontology Module Entries

  • The four modules encode task-specific semantics, including geochemical aliases and detection-limit rules, assay vocabularies, composition units, and linked ML result attributes.The geochemistry examples distinguish about 90 deposit keyword mappings from a 189-type classifier vocabulary and map more than 50 method variants to about 12 canonical methods.Source: §E.1 · p. 19, §E.2 · p. 19, §E.3 · p. 19

F. Prompt Templates

  • Prompt conventions can introduce systematic interpretation errors: the ChemTables prompt omits CC50 despite its module support, and composition totals cannot identify mol% versus wt%.GeoChem open backbones use a 512-token cap, but all three deposit classifiers still return malformed JSON; the pipeline sets deposit_type=null.Source: F. Prompt Templates · pp. 21–22
  • Three exemplars are motivated for ChemTables only; the GeoChem few-shot baseline lacks supplementary spreadsheets and uses a compact per-row schema.Thus its comparison with ArticleMiner includes input-access and representation differences; the exemplar-count analysis does not establish context sufficiency elsewhere.Source: F. Prompt Templates · p. 22

G. Sonnet 4.6 Sensitivity on PDF-only Domains

  • On small PDF-only ChemTables and MLTables samples, component effects differ by backbone, and paper-intelligence removal changes direction between Haiku and Sonnet.The ablations are task- and backbone-specific diagnostics, not a universal component ranking.Source: G. Sonnet 4.6 Sensitivity on PDF-only… · p. 22

H. Extended Results

Extended analyses show that component effects vary by task and backbone, while GeoChem tier scores and open-model results have important scope qualifications. Several appendix comparisons are exploratory or use different configurations and subsets.

H. Extended Results

  • The protocol-component sequence is historical rather than a controlled one-factor ablation, and backbone-output overlap measures agreement rather than correctness.Table 8’s changes in T4 cannot be assigned solely to the deposit classifier.Source: H. Extended Results · p. 22, Table 8 · p. 23, Table 9 · p. 23
  • Removing paper intelligence raises ChemTables F1 by 12.5 under Sonnet, but the effect is backbone-dependent and does not establish its mechanism.The appendix describes the associated precision increase as an interaction, not a demonstrated mechanism.Source: H. Extended Results · p. 23
  • Vision produces the largest losses under Sonnet, whereas validation removal produces larger losses under Haiku; GeoChem’s supplementary-data path makes LLM-side components less influential.The appendix cautions that neither grid establishes universal component effects.Source: H. Extended Results · p. 23

H.1. Full cross-model grid

  • Tables 10–13 provide precision, recall, and F1 grids across eight backbones and both input modalities; Figure 3’s strict tuple-F1 and GeoChem Overall S are distinct metrics.The grid’s ChemTables benchmark contains nine publications; its caption identifies raw-PDF and pre-parsed inputs.Source: Figure 3 · p. 23, §H.1 · p. 24, Table 10 · p. 24

H.2. GeoChem per-paper 4-tier breakdown (𝑛=28)

  • GeoChem’s 28-paper headline configuration has mean Overall S of 76.0, with wide T2 and T3 intervals and a long left tail across papers.The 95% confidence interval for Overall S is 70.1–81.4; a small number of layout-only or borderless-table papers score near zero.Source: §H.2 · pp. 24–25

H.3. Interpretation of component and backend analyses

  • Backend comparisons use a corrected-ground-truth 26-paper subset, not the 28-paper headline reference; removing Camelot changes results differently across the three external benchmarks.The appendix warns against reading component-removal and backend panels as one controlled sequence.Source: §H.3 · p. 26

H.4. Open-LLM GeoChem 4-tier breakdown (𝑛=26, max-tokens=512)

  • On the 26-paper GeoChem subset, open-backbone Overall scores range from 52.6 to 54.0, while malformed deposit-classifier JSON under the 512-token cap confounds comparison with closed models.The reported T2, T3, and T4 aggregates are identical across the three open runs, consistent with their shared supplementary-file parser, but this does not imply identical cell-level outputs.Source: §H.4 · p. 27

I. Hardware, Software, and Reproducibility

The paper reports hardware, software families, and inference settings, but the released materials do not fully lock the environment or map every historical result to its configuration.

I. Hardware, Software, and Reproducibility

  • Experiments used Ubuntu 22.04, CUDA 12.4, an NVIDIA H100 80GB GPU, and vLLM 0.6.x for open-model inference.Closed models were called through Anthropic, OpenAI, and Google SDKs in April–May 2026.Source: §I · p. 27
  • The reported model identifiers and parser versions are specified, and temperature was zero with resampling and exemplar-ordering seed 42.MinerU’s OCR path was disabled when its weights were unavailable.Source: §I · p. 27
  • Software versions are families rather than a complete executable lock, and the release lacks a complete manifest linking every historical result to its configuration.Reproduction requires corresponding settings and output files; running the latest default pipeline is insufficient.Source: §I · p. 27

J. Knowledge-Graph Output: Export Structure

The exporter represents recovered samples, measurements, and source metadata in JSON-LD, while GeoChem ground truth follows a documented expert-curation protocol. Tuple-level evaluation does not establish downstream RDF interoperability or correctness.

J. Knowledge-Graph Output: Export Structure

  • The JSON-LD exporter creates one node per sample row, with a paper-scoped normalized identifier, sample name, and optional mineral, deposit, and analytical-method fields.Each nonempty elemental value becomes a measurement with element and value_ppm, plus a unit when available.Source: J. Knowledge-Graph Output: Export… · p. 27
  • Provenance records the source PDF stem and pipeline model, but do not provide a complete PROV mapping or cell-level evidence locator.The implemented representation does not establish verified external ontology alignment, native equivalent Turtle output, or downstream query correctness.Source: J. Knowledge-Graph Output: Export… · p. 28

K. GeoChem-28 Annotation Protocol

  • GeoChem-28 covers 28 peer-reviewed papers on LA-ICP-MS trace-element analyses of sulfide and sulphosalt minerals, with annotations checked by a second annotator and reconciled against source files.The benchmark uses the 209-column CMiO-MIN schema, an extension of CMiO/Hofstra.Source: K. GeoChem-28 Annotation Protocol · p. 28
  • Each GeoChem row represents one sample–mineral analysis; units remain as reported, while below-detection values use −L or −99999 and missing or unmeasured entries remain blank.Reference-material rows are retained, and two data-reuse entries lack standalone PDFs, reducing the open-LLM breakdown to 26 papers.Source: K. GeoChem-28 Annotation Protocol · pp. 28–29
  • GeoChem’s ground-truth annotations are released under CC-BY-4.0, but this license does not grant redistribution rights to source-paper PDFs.Source: K. GeoChem-28 Annotation Protocol · p. 29

Related papers

Loading 2609.25607v1…