Pith. sign in

REVIEW 3 major objections 4 minor 39 references

A hybrid rule-LLM pipeline extracts Scope 1, 2, and 3 greenhouse-gas disclosures from sustainability reports with ~99% F1, anchoring every value to a page of source evidence.

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

2026-08-01 18:56 UTC pith:SOCA3RDJ

load-bearing objection Useful evidence-grounded extraction pipeline, but the headline F1 is an upper bound under the gold-set exclusions and the prediction benchmark is circular. the 3 major comments →

arxiv 2607.17122 v1 pith:SOCA3RDJ submitted 2026-07-19 cs.CL

Scope3Trace: Evidence-Based Identification and Extraction of Scope 3 GHG Emissions from Sustainability Reports

Siyuan Zheng , Yifan Duan , Chao Xue , Flora D. Salim This is my paper
classification cs.CL
keywords Scope 3 emissionsgreenhouse gas accountinginformation extractionsustainability reportsLLM pipelineevidence groundingESG datamultimodal dataset
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.

Scope3Trace aims to show that Scope 3 greenhouse-gas disclosures - the majority of most corporate carbon footprints and the least standardized part of carbon accounting - can be extracted from real-world sustainability reports automatically, accurately, and with evidence a skeptic can check. The paper's central claim is that a hybrid design, in which an LLM only locates relevant pages, classifies content, and reconstructs tables while deterministic rules handle numeric parsing, category mapping, and evidence verification, reaches ~0.99 F1 on Scope 1/2/3 totals and 0.97 on category-level extraction. The framework also produces a dual-level dataset (organization and building) with provenance labels that keep reported values distinct from modeled or estimated ones. If this claim holds, large-scale, auditable Scope 3 analysis no longer depends on expensive per-document manual annotation.

Core claim

Scope3Trace's central claim is that Scope 3 emission figures, widely considered the least standardized and most opaque part of corporate carbon accounting, can be pulled from real-world sustainability reports with high accuracy when the language model is kept on a short leash. The pipeline parses PDFs with OCR, uses an LLM to locate emission-related pages and reconstruct tables, and then hands every numeric value to deterministic rules for unit normalization, category mapping, and evidence checking. Each extracted value is stored with the page number and a verbatim text snippet that supports it. On a manually annotated gold set of 200 reports, the paper reports Scope 1/2/3 F1 of 0.99/0.98/0.

What carries the argument

The load-bearing mechanism is the bounded hybrid design: the LLM performs only semantic tasks (page-level localization of emissions content, sentence classification, table interpretation and reconstruction), while rule-based components perform deterministic tasks (numeric parsing, unit normalization, GHG-category mapping, span detection, and evidence verification against OCR text). The central object is the evidence link - a (page, verbatim snippet) pair attached to every extracted value. Ablations show that removing table reconstruction or evidence grounding costs the most accuracy, identifying those two components as the source of the gains.

Load-bearing premise

The gold evaluation set was annotated by the authors under strict exclusion criteria - no offsets, credits, intensity metrics, combined totals, or percentage-only disclosures - so the reported F1 numbers describe a curated subset of disclosures; if those exclusions remove most of the genuinely hard extraction cases, performance on arbitrary reports would be lower.

What would settle it

Have an independent team annotate a random sample of uncurated sustainability reports - including offsets, intensity metrics, percentage-only disclosures, and combined totals - run Scope3Trace on the same PDFs, and compare value-level F1; a material drop below the reported 0.99/0.98/0.99 across Scope 1/2/3 would refute the claim that the framework generalizes beyond the gold-set conditions.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Scope 3 extraction can scale beyond hand-annotated samples: the same hybrid pipeline processes a full PDF for roughly $0.385 per document with no per-document human verification.
  • Every value's page-level evidence makes the output auditable by a third party, moving ESG data from document-level references to value-level provenance.
  • Because ablations show table reconstruction and evidence grounding are the main accuracy drivers, future systems can focus effort on those components rather than on prompt engineering.
  • The dual-level dataset supports building-level attribution: among the owning companies of the ~580 building records with directly reported per-building Scope 3 disclosures, 96% are linked to corresponding organization-level disclosures.
  • Building-level Scope 3 prediction is feasible from the dataset's features (R2 = 0.46 for total Scope 3, macro R2 = 0.54 for categories), so the data supports modeling as well as extraction.

Where Pith is reading between the lines

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

  • This suggests the pipeline could be repurposed as an automated audit tool: regulators could run it on submitted reports and check whether cited evidence pages actually contain the disclosed values, turning extraction into a verifier of disclosure quality.
  • The same evidence links form a natural benchmark for LLM grounding: a model's extraction is correct only if the cited page contains the value, so the dataset could be used to measure hallucination rates in document-understanding tasks.
  • A stress test that re-introduces the excluded item types (offsets, intensity metrics, percentage-only disclosures) would reveal whether the near-ceiling accuracy is a property of the pipeline or of the simplified gold-set task.
  • The reported and estimated layers could be combined to compute coverage ratios - for example, summed building-level estimates versus organization totals - exposing boundary gaps in a way the paper's diagnostics only begin to quantify.

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

3 major / 4 minor

Summary. The paper presents Scope3Trace, an evidence-grounded information extraction framework for Scope 1/2/3 GHG emissions from corporate sustainability reports. The pipeline combines PDF/OCR parsing, LLM-based page localization and table reconstruction, and hybrid rule-based/LLM extraction with page-level evidence anchoring. Using this framework, the authors construct a dual-level dataset with organization-level disclosures and building-level records (54,361 building-year records, mostly Australia/Europe). They report near-perfect extraction performance (Scope1/2/3 F1 0.98–0.99, Category F1 0.97, EM 0.96) on a 200-document gold evaluation set, compare against LLM-only, RAG, and multi-agent baselines, and present a downstream building-level Scope 3 prediction benchmark (R²=0.46 total, macro R²=0.54 category).

Significance. If the extraction results hold across real-world report heterogeneity, the framework would be a valuable contribution: it ships code, uses deterministic rule components to bound LLM behavior, provides page-level evidence links, and demonstrates robustness across LLM backbones with a careful cost analysis. The dataset, with explicit provenance labels and physical consistency checks, is a useful resource for sustainability reporting research. However, the significance is tempered by two load-bearing evaluation concerns: the gold set excludes several common disclosure formats, and the downstream prediction benchmark is partly circular. Both are fixable, but they currently weaken the paper's headline claims.

major comments (3)
  1. [Appendix G.4, Table 1] The gold evaluation set explicitly excludes carbon offsets/RECs, intensity metrics, reduction targets/baselines, percentage-only disclosures, and combined totals (e.g., Scope 1+2). These are common in real sustainability reports, so the 0.96–0.99 F1/EM in Table 1 is computed only for clean absolute totals and category lines. The paper's motivation emphasizes heterogeneity (§3.1) and claims 'audit-grade extraction' (Figure 2); the evaluation does not sample the heterogeneity that motivated the framework. Please report performance on the excluded item types or explicitly reframe the accuracy claim as limited to standalone absolute disclosures.
  2. [§4.3, Appendix C.2/C.3, Appendix N.1 (Table 14)] The building-level prediction benchmark is partly circular. Category 2 and Category 13 targets are defined in Appendix C.2 and C.3 as deterministic functions of FloorArea and electricity consumption: Cat2 = FloorArea × EF_embodied / 1000 and Cat13 = TenantElectricity × EF_grid. Both floor_area and energy/tenant_electricity variables appear in the 'full features' input set (Table 14). Thus the R²=0.46 (total) and macro R²=0.54 (category) largely reflect the model's ability to invert the derivation formulas rather than the dataset's predictive utility for unseen independent data. The utility claim would be defensible if these direct inputs were removed from the feature set, or if the target variables came from independently reported values rather than the same activity data used as features.
  3. [§G.1, §G.7] The gold annotations were produced by co-authors with 83% double-annotation and κ=0.89. While numeric values are verifiable, the strict exclusion criteria (§G.4) and consensus resolution of disagreements may jointly bias the gold set toward extraction-friendly disclosures. Please provide a per-format breakdown (e.g., F1 for paragraph vs. table vs. footnote disclosures) and a detailed account of disagreement resolution. This would help readers judge whether the 'near-perfect' numbers are an artifact of a simplified task definition.
minor comments (4)
  1. [§3.1] The sentence 'including scanned PDFs, formatted text, and mixed text-image layouts' follows an incomplete clause ('with inconsistent wording... varying document formats. including...'). Please rephrase for grammar and clarity.
  2. [Figure 2] The label 'audit-grade Scope 1/2/3 emissions data construction' overstates the verification. No external audit was performed; the pipeline provides evidence links and confidence scores. Consider 'evidence-graded' or 'traceable construction' to avoid implying third-party assurance.
  3. [Table 1 caption] The caption mentions 'a retrieval-augmented LLM pipeline' but no such row appears in the table. Either add the row or remove the description.
  4. [Appendix E] The sentence 'The results are reported as mean ± standard deviation over three runs' appears twice near the table. Consider consolidating to avoid redundancy.

Circularity Check

1 steps flagged

Building-level prediction benchmark is partly circular: Cat2/Cat13 targets are defined from the same floor-area/electricity features the model is given.

specific steps
  1. fitted input called prediction [Appendix C.2–C.3 (definitions); Appendix N.1 and Tables 12–13 (prediction benchmark); M.1.2]
    "Scope 3 category 2 emissions at the building level are approximated using floor-area-based embodied carbon intensity factor: Cat2tCO2e = Floor Aream2 × EFembodied / 1000 ... Cat13tCO2e = Tenant Electricity kW h × EFgrid, where Tenant Electricity kW h is estimated from total building electricity consumption scaled by the commercial buildings proportion ... Full features: the complete Scope3Trace tabular schema enumerated in Appendix N.1."

    The prediction targets are the very quantities the pipeline constructs from the regressors. Cat2 is defined as floor area × embodied-carbon factor; Cat13 as tenant electricity × grid factor. Appendix N.1's full feature set includes floor_area, energy_electricity_kwh, tenant_electricity_kwh, p_com/p_res, NABERS intensities, and Scope 2 emissions, so a fitted regressor can approximately recover the defining formulas. Table 11 shows 99.99% of Cat2 and 100% of Cat13 values are estimated/modeled, and the Limitations section states these are 'structured approximations, not ground-truth measurements.' The R2=0.46 total and macro R2=0.54 category results in Tables 12/13 therefore partly re-learn the construction formulas rather than measuring independent predictive signal. The extraction results i

full rationale

Scope3Trace's central extraction claim is self-contained: Table 1 compares against rule-based, LLM-only, RAG, and multi-agent baselines on a manually annotated 200-report gold set with page-level evidence and inter-annotator agreement κ=0.89. The extraction pipeline's high F1 does not reduce to its inputs, and no load-bearing self-citations appear. The circularity is confined to the downstream 'dataset utility' prediction benchmark: the building-level Cat2/Cat13 targets are defined in Appendix C from floor area and electricity, which are themselves listed in Appendix N.1 as full input features, and Table 11 shows those targets are almost entirely estimated/modeled rather than reported. Thus the R2 values in Tables 12–13 partly measure the model recovering the construction formulas. I also examined the Appendix G.4 gold-set exclusions (offsets, intensity metrics, reduction targets, combined totals, percentage-only disclosures): these are a benchmark-scope/generalization limitation, not a circular reduction—they do not make the extraction scores implied by the definitions—so they are noted but not counted in the score. Overall partial circularity in the prediction benchmark gives score 6; the extraction core is independent.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The extraction pipeline assumes LLM semantic judgments at temperature=0 are reliable and that deterministic rules cover unit and category variations. The building-level layer assumes external emission factors and public inventories are valid for the assets. The prediction benchmark assumes the constructed target values are meaningful independent outcomes, which is questionable because Cat2 and Cat13 are functions of the input features.

free parameters (2)
  • EF_grid = region/year-specific, from public databases
    Electricity grid emission factor (tCO2e/kWh) used in C.1 and C.3 to derive Scope 2 and Category 13 emissions. It is externally sourced, not fitted to the paper's data, but it is a chosen parameter that determines derived building values.
  • EF_embodied = per building typology, from public databases
    Embodied carbon intensity (tCO2e/m2) used in C.2 to derive Category 2 emissions. The paper selects it by typology, and the resulting values are labeled as modeled/estimated rather than reported.
axioms (5)
  • domain assumption GHG Protocol Scope 1/2/3 definitions and 15 Scope 3 categories are accepted as ground truth.
    The schema and category mapping in Sections 3 and A.1 depend on the GHG Protocol standard without proving it.
  • domain assumption OCR output faithfully represents the source PDFs, including scanned and mixed-layout reports.
    The pipeline converts PDFs to text via OCR (Section 3.3.2) and evidence grounding links values to OCR-derived snippets; OCR errors would propagate into extraction and evidence.
  • domain assumption Public emission factor and building inventory databases are authoritative for derived building-level values.
    Appendices C and H use NABERS/NGER, grid factors, and embodied-carbon databases as trusted sources for estimation.
  • domain assumption The gold annotation protocol with strict exclusions captures the true extraction target without bias.
    Appendix G states annotations were produced by co-authors and exclusion criteria remove offsets, intensities, combined totals, and percentage-only disclosures; if these exclusions make the task easier, the reported F1 is optimistic.
  • domain assumption All processed disclosures are in English.
    Appendix A states all source disclosures and extracted records are in English, limiting generalization to non-English reports.

pith-pipeline@v1.3.0-alltime-deepseek · 20757 in / 10186 out tokens · 100429 ms · 2026-08-01T18:56:01.395983+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Scope3Trace: Evidence-Based Identification and Extraction of Scope 3 GHG Emissions from Sustainability Reports." pith.science (2026). https://pith.science/paper/SOCA3RDJ

@misc{pith2026260717122,
  author       = {Pith},
  title        = {Pith review of: Scope3Trace: Evidence-Based Identification and Extraction of Scope 3 GHG Emissions from Sustainability Reports},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SOCA3RDJ}},
  note         = {Machine review of arXiv:2607.17122}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Scope 3 greenhouse gas (GHG) emissions account for the majority of corporate carbon footprints, yet remain difficult to analyze at scale due to sparse disclosures, heterogeneous report document formats, and limited evidence traceability. Existing approaches typically rely on large language models to extract emissions information from ESG reports, but often lack explicit evidence grounding or depend on costly manual annotation and verification to ensure extraction reliability. To address these challenges, we propose Scope3Trace, an evidence-grounded information extraction framework designed to extract interpretable and traceable Scope 3 emissions information from real-world ESG and sustainability reports. The framework integrates a document information extraction pipeline that performs PDF collection and OCR parsing, LLM-assisted page localization and table reconstruction, and hybrid rule-LLM extraction of organization- and building-level emissions disclosures with evidence-grounded verification. Building upon this framework, we further contribute a dual-level, evidence-grounded, multimodal dataset comprising organization-level Scope 3 disclosures extracted from heterogeneous sustainability reports. Scope3Trace enables reliable extraction and transparent integration of heterogeneous sustainability disclosures, achieving high accuracy in extracting Scope 1-3 totals and category-level disclosures from sustainability reports.

Figures

Figures reproduced from arXiv: 2607.17122 by Chao Xue, Flora D. Salim, Siyuan Zheng, Yifan Duan.

Figure 1
Figure 1. Figure 1: Illustration of Scope 1,2,3 emissions under [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the Scope3Trace multi-agent system for dataset construction. The pipeline is orchestrated [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Geographic coverage of the Scope3Trace dataset. records across multiple countries in Europe, Aus￾tralia, and Asia, spanning reporting years 2011– 2026, as illustrated in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Building-level availability of major Scope 3 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustrative example of a gold annotation [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Data provenance distribution (non-residential [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Value provenance and confidence summary for non-residential buildings in Scope3Trace. industry reporting practices where building-level Scope 3 disclosure remains sparse. J.2 Provenance Metadata Completeness Scope3Trace explicitly records provenance meta￾data including estimation method, data source, and quality indicators [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Confidence distribution by Scope categories [PITH_FULL_IMAGE:figures/full_fig_p019_9.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

39 extracted references · 4 canonical work pages

  1. [2]

    Jacob Beck, Anna Steinberg, Andreas Dimmelmeier, Laia Domenech Burin, Emily Kormanyos, Maurice Fehr, and Malte Schierholz. 2025. https://doi.org/10.1038/s41597-025-05664-8 Addressing data gaps in sustainability reporting: A benchmark dataset for greenhouse gas emission extraction . Scientific Data, 12:1497

    doi:10.1038/s41597-025-05664-8 2025

  2. [3]

    Andreas Dimmelmeier, Hendrik Doll, Malte Schierholz, Emily Kormanyos, Maurice Fehr, Bolei Ma, Jacob Beck, Alexander Fraser, and Frauke Kreuter. 2024. https://doi.org/10.18653/v1/2024.climatenlp-1.2 Informing climate risk analysis using textual information - a research agenda . In Proceedings of the 1st Workshop on Natural Language Processing Meets Climate...

    work page doi:10.18653/v1/2024.climatenlp-1.2 2024

  3. [4]

    Yi Ding, Xushuo Tang, Zhengyi Yang, Wenqian Zhang, Simin Wu, Yuxin Huang, Lingjing Lan, Weiyuan Li, Yin Chen, Mingchen Ju, Wenke Yang, Thong Hoang, Mykhailo Klymenko, Xiwei Zu, and Wenjie Zhang. 2025. https://arxiv.org/abs/2511.21712 Euleresg: Automating esg disclosure analysis with llms . arXiv preprint

    arXiv 2025

  4. [5]

    Andrew Dumit, Krishna Rao, Travis Kwee, Varsha Gopalakrishnan, Katherine Tsai, and Sangwon Suh. 2024. https://www.climatechange.ai/papers/neurips2024/70 Atlas: A spend classification benchmark for estimating scope 3 carbon emissions . In NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning

    2024

  5. [7]

    Greenhouse Gas Protocol . 2011. https://ghgprotocol.org Corporate value chain (scope 3) accounting and reporting standard

    2011

  6. [8]

    Greenhouse Gas Protocol . 2019. You too can master value chain emissions. https://ghgprotocol.org/blog/you-too-can-master-value-chain-emissions

    2019

  7. [9]

    Yanming Guo, Jin Ma, Xiao Qian, and Kevin Credit. 2025. https://huggingface.co/datasets/Yvnminc/ExioNAICS Exionaics: Enterprise-level emission estimation dataset with large language models . In Proceedings of the ICLR 2025 Workshop on Tackling Climate Change with Machine Learning

    2025

  8. [10]

    Hertwich and Richard Wood

    Edgar G. Hertwich and Richard Wood. 2018. https://doi.org/10.1088/1748-9326/aae19a The growing importance of scope 3 greenhouse gas emissions from industry . Environmental Research Letters, 13(10):104013

    work page doi:10.1088/1748-9326/aae19a 2018

  9. [11]

    Manfred Lenzen, Daniel Moran, Keiichiro Kanemoto, Barney Foran, Lara Lobefaro, and Arne Geschke. 2012. https://doi.org/10.1038/nature11145 International trade drives biodiversity threats in developing nations . Nature, 486(7401):109--112

    work page doi:10.1038/nature11145 2012

  10. [12]

    M. Li. 2025. Greenhouse gas performance and disclosure: new global evidence. Energy Economics

    2025

  11. [13]

    N. H. Luu. 2025. Does mandatory greenhouse gas emissions reporting reduce greenwashing? Journal of Environmental Management

    2025

  12. [14]

    Lokesh Mishra, Sohayl Dhibi, Yusik Kim, Cesar Berrospi Ramis, Shubham Gupta, Michele Dolfi, and Peter Staar. 2024. https://aclanthology.org/2024.climatenlp-1.15/ Statements: Universal information extraction from tables with large language models for esg kpis . In Proceedings of the 1st Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2024)

    2024

  13. [15]

    Jeff Sternberg. 2022. https://cloud.google.com/blog/topics/sustainability/cop27-climate-risk-analysis-and-esg-entity-extraction At finance day, exploring climate risk analysis and esg entity extraction . Google Cloud Blog, Sustainability Topics

    2022

  14. [21]

    Scientific Data , volume =

    Jacob Beck and Anna Steinberg and Andreas Dimmelmeier and Laia Domenech Burin and Emily Kormanyos and Maurice Fehr and Malte Schierholz , title =. Scientific Data , volume =. 2025 , publisher =

    2025

  15. [22]

    Proceedings of the ICLR 2025 Workshop on Tackling Climate Change with Machine Learning , year =

    Yanming Guo and Jin Ma and Xiao Qian and Kevin Credit , title =. Proceedings of the ICLR 2025 Workshop on Tackling Climate Change with Machine Learning , year =

    2025

  16. [23]

    arXiv preprint arXiv:2312.17264 , year =

    Yi Zou and Mengying Shi and Zhongjie Chen and Zhu Deng and ZongXiong Lei and Zihan Zeng and Shiming Yang and HongXiang Tong and Lei Xiao and Wenwen Zhou , title =. arXiv preprint arXiv:2312.17264 , year =

    Pith/arXiv arXiv

  17. [24]

    Proceedings of the 1st Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2024) , year =

    Lokesh Mishra and Sohayl Dhibi and Yusik Kim and Cesar Berrospi Ramis and Shubham Gupta and Michele Dolfi and Peter Staar , title =. Proceedings of the 1st Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2024) , year =

    2024

  18. [25]

    arXiv preprint , year =

    EulerESG: Automating ESG Disclosure Analysis with LLMs , author =. arXiv preprint , year =

  19. [26]

    Hertwich and Richard Wood , title =

    Edgar G. Hertwich and Richard Wood , title =. Environmental Research Letters , volume =. 2018 , doi =

    2018

  20. [27]

    NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning , year =

    Andrew Dumit and Krishna Rao and Travis Kwee and Varsha Gopalakrishnan and Katherine Tsai and Sangwon Suh , title =. NeurIPS 2024 Workshop on Tackling Climate Change with Machine Learning , year =

    2024

  21. [28]

    arXiv preprint arXiv:1908.10084 , year =

    Nils Reimers and Iryna Gurevych , title =. arXiv preprint arXiv:1908.10084 , year =

    Pith/arXiv arXiv 1908

  22. [29]

    2022 , month = nov, url =

    Sternberg, Jeff , title =. 2022 , month = nov, url =

    2022

  23. [30]

    Proceedings of the 1st Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2024) , month =

    Informing climate risk analysis using textual information - A research agenda , author =. Proceedings of the 1st Workshop on Natural Language Processing Meets Climate Change (ClimateNLP 2024) , month =. 2024 , address =

    2024

  24. [31]

    Proceedings of the ClimateNLP Workshop , year =

    Webersinke, Nicolas and Kraus, Mathias and Bingler, Julia and Leippold, Markus , title =. Proceedings of the ClimateNLP Workshop , year =

  25. [32]

    ICML Workshop on Tackling Climate Change with Machine Learning , year =

    Luccioni, Alexandra Sasha and Schmidt, Victor and Vardanyan, Varun , title =. ICML Workshop on Tackling Climate Change with Machine Learning , year =

  26. [33]

    arXiv preprint arXiv:2109.08901 , year =

    Gupta, Rachit and Modgil, Saurabh and Jha, Radhika , title =. arXiv preprint arXiv:2109.08901 , year =

    Pith/arXiv arXiv

  27. [34]

    arXiv preprint arXiv:2310.05628 , year =

    Bronzini, Francesco and others , title =. arXiv preprint arXiv:2310.05628 , year =

    Pith/arXiv arXiv

  28. [35]

    Nature Sustainability , year =

    Winston Yap and Abraham Noah Wu and Clayton Miller and Filip Biljecki , title =. Nature Sustainability , year =. doi:10.1038/s41893-025-01615-8 , note =

    doi:10.1038/s41893-025-01615-8

  29. [36]

    arXiv preprint arXiv:2405.14135 , year=

    Space-aware Socioeconomic Indicator Inference with Heterogeneous Graphs , author=. arXiv preprint arXiv:2405.14135 , year=

    Pith/arXiv arXiv

  30. [37]

    Corporate Value Chain (Scope 3) Accounting and Reporting Standard , year =

  31. [38]

    arXiv preprint arXiv:2305.11906 , year=

    Abominable greenhouse gas bookkeeping casts serious doubts on climate intentions of oil and gas companies , author=. arXiv preprint arXiv:2305.11906 , year=

    Pith/arXiv arXiv

  32. [39]

    Nature , volume =

    Lenzen, Manfred and Moran, Daniel and Kanemoto, Keiichiro and Foran, Barney and Lobefaro, Lara and Geschke, Arne , title =. Nature , volume =. 2012 , doi =

    2012

  33. [40]

    Evaluating Corporate Carbon Emissions Reporting: Assessing Transparency and Completeness with the Carbon Integrity Index , journal =

    Traub, Jos. Evaluating Corporate Carbon Emissions Reporting: Assessing Transparency and Completeness with the Carbon Integrity Index , journal =. 2025 , volume =. doi:10.3390/su17177628 , url =

    work page doi:10.3390/su17177628 2025

  34. [41]

    2019 , howpublished=

    You Too Can Master Value Chain Emissions , author=. 2019 , howpublished=

    2019

  35. [42]

    arXiv preprint arXiv:2312.10997 , volume=

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

    Pith/arXiv arXiv

  36. [43]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  37. [44]

    Journal of Environmental Management , year=

    Does mandatory greenhouse gas emissions reporting reduce greenwashing? , author=. Journal of Environmental Management , year=

  38. [45]

    Energy Economics , year=

    Greenhouse gas performance and disclosure: new global evidence , author=. Energy Economics , year=

  39. [46]

    arXiv preprint arXiv:2308.08155 , year=

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation , author=. arXiv preprint arXiv:2308.08155 , year=

    Pith/arXiv arXiv