Pith. sign in

REVIEW 4 major objections 5 minor 41 references

Measuring Corporate Human Capital Disclosures: Lexicon, Data, Code, and Research Opportunities

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A machine-learning word model trained on 3,953 corporate 10-K human capital sections produces a 1,285-term dictionary for measuring workforce disclosure in any corporate text.

desk verdict Useful, honest measurement resource; 'high construct validity' is asserted rather than demonstrated, so accept with revisions requiring a gold-standard benchmark. read the letter →

arxiv 2506.10155 v1 pith:JEQ6A7X5 submitted 2025-06-11 cs.CL cs.AIcs.LG

Elizabeth Demers , Victor Xiaoqi Wang , Kean Wu This is my paper · ORCID
classification cs.CLcs.AIcs.LG
keywords humancapitaldisclosureword2veclexiconconstructionSEC2020mandateBERTfine-tuning10-Kfilingstextualanalysiscorporatemeasurement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Human capital is widely described as a company's most important asset, but it has no standard measurement or disclosure rule. This paper tries to fix the measurement problem by building an extensive lexicon of 1,285 human-capital terms, trained with the word2vec algorithm on the human-capital disclosures that 3,953 firms filed under the SEC's 2020 mandate. Because the training corpus is the actual HC-specific section of 10-Ks, the authors argue the resulting word list measures what it claims to measure more directly than earlier keyword lists or a model trained on entire 10-Ks. They release the lexicon, the disclosure texts, and Python code, and show the data can also fine-tune a BERT model to identify HC sentences. If the approach works, researchers can measure how firms talk about employees across proxy statements, quarterly reports, conference calls, and sustainability reports, at a scale manual coding cannot reach.

What carries the argument

The machinery is word2vec, a neural word-embedding model that represents each word by a numeric vector learned from the surrounding words and treats words used in similar contexts as similar. Trained on the confirmed HC disclosures, it expands the seed-word list by vector similarity; a cosine-similarity cutoff of 0.5 in absolute value plus manual screening converts the vector space into a 1,285-term dictionary. The same hand-collected HC corpus doubles as labeled data for fine-tuning a base BERT model, a pre-trained transformer language model, into a binary HC classifier, giving the paper's measurement toolkit a second, machine-learning-based mode.

What would settle it

Apply the released lexicon, with the paper's own counting rules, to a corpus of clearly non-human-capital text, such as routine risk-factor paragraphs or financial footnotes, and compare keyword density with the density in confirmed HC disclosures; alternatively, have independent coders judge a random sample of sentences flagged by the lexicon in proxy statements. If non-HC text matches at a rate close to HC text, or if a large share of flagged sentences is judged unrelated to human capital, the construct-validity claim is refuted.

Watch

Extended reading notes

Core claim

The paper's central claim is that a word2vec model trained on a corpus that is unambiguously about human capital can yield an extensive, multidimensional dictionary of contemporary HC language. The training corpus is hand-collected from the human capital sections of 3,953 unique 10-K filings from the first year of the SEC's 2020 disclosure mandate, about two million words. Starting from seed words drawn from standard-setter frameworks and manual reading, the authors keep words with cosine similarity at least 0.5 to any seed word, along with highly negative similarities to catch antonyms, manually screen 7,018 candidates, and end with 1,285 terms in five categories: DEI; health and safety; labor relations and culture; compensation and benefits; and demographics and other. They report that applying the lexicon to proxy statements from 1994 to 2022 picks up known regulatory shocks, such as the 2007 compensation disclosure expansion, which they take as evidence that the measure is capturing real disclosure behavior. They also fine-tune a base BERT model on 83,961 HC sentences and report 97.46 percent accuracy on held-out 10-K sentences and an F1 of about 85 percent on manually labeled 10-Q sentences.

Load-bearing premise

The load-bearing premise is that the hand-collected human capital sections of the 10-K filings really are about human capital management, so the word associations learned from them will transfer to other corporate documents; if those sections are padded with generic business, legal, or risk boilerplate, the resulting dictionary will match too many non-human-capital texts.

Editorial extensions

If this is right

  • A researcher can count the 1,285 terms in any set of corporate texts—proxy statements, 10-Qs, conference call transcripts, sustainability reports—to produce firm-level human capital disclosure measures without manual coding.
  • Separate lists for DEI, health and safety, labor relations and culture, compensation and benefits, and demographics allow studies to ask which dimension of human capital management a firm is discussing.
  • The released corpus of roughly two million words of confirmed HC disclosure text can be used to fine-tune BERT or another language model; the paper's own fine-tuned BERT reaches about 97 percent accuracy on held-out 10-K sentences and an F1 near 85 percent on 10-Q sentences.
  • Supplying a new set of seed words to the released code should produce a comparable dictionary for another construct, so the data pipeline is reusable beyond human capital.

Reading between the lines

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

  • Because the lexicon is built from 2020–2021 disclosures, its COVID-heavy health and safety terms will age; users working on earlier or later periods may need to refresh the seed words, something the authors do not claim to do.
  • Counting HC keywords alongside independently measured workforce outcomes such as turnover, injury rates, or pay gaps could turn the lexicon into a screen for social washing, a use the paper lists as an open question rather than a demonstrated result.
  • The labeled corpus could support multi-label classification into the five HC categories, or extraction of quantitative HC metrics, going beyond the binary HC-versus-not classification the paper demonstrates.
  • A formal precision study on non-HC text would quantify the false-positive rate of the lexicon; the proxy-statement application is suggestive but does not measure precision directly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper develops a human capital (HC) disclosure lexicon using word2vec trained on hand-collected, machine-readable HC disclosures from 3,953 unique 10-K filings in the first year of the SEC's 2020 HC disclosure rule (November 2020 to November 2021). Starting from seed words in five categories (DEI; health and safety; labor relations and culture; compensation and benefits; demographics and other), the authors retain words with cosine similarity at or above 0.5 (or at or below -0.5), manually screen 7,018 candidates down to 1,285 terms, and categorize them into five panels plus a COVID-19 subcategory. The paper releases the lexicon, the corpus of HC disclosures, and Python code, and illustrates two uses: measuring HC keyword trends in proxy statements over 1994-2022, and fine-tuning BERT for binary HC-sentence classification with reported accuracy, precision, recall, and F1 metrics. It concludes with research opportunities for accounting and finance scholars.

Significance. If the lexicon has the claimed construct validity, this is a useful public resource for measuring corporate HC disclosure at scale in 10-Ks, proxy statements, conference calls, and other texts, and the released code and data would support replication and extension to other constructs. The paper is transparent in sharing the lexicon, the labeled corpus, and code, which is a concrete strength that lowers the barrier for subsequent research. The BERT fine-tuning illustration also demonstrates a plausible complementary approach. However, the central claim of 'high construct validity' is asserted rather than quantitatively established: the manual screening step has no reliability check, and the proxy-statement validation is a face-validity trend analysis rather than a benchmark against human-coded text or existing lexicons. The significance of the contribution is therefore conditional on additional validation.

major comments (4)
  1. [Section III (Final Word lists)] The manual screening step that reduces the 7,018 candidate terms to the final 1,285 terms is load-bearing for the paper's central claim of high construct validity, but no inter-rater reliability, coding protocol, or resolution procedure is reported. Because the final lexicon is the primary deliverable, the authors should provide reliability evidence (e.g., dual coding of a random sample with Cohen's kappa) and a transparent exclusion rule, or temper the construct-validity claim accordingly.
  2. [Section IV (An Application of Our Lexicon; Figures 5 and 6)] The proxy-statement illustration is too weak to support the strong claim of high construct validity. The sharp increases in 2007 and 2010-11 are concentrated in Compensation and Benefits and Demographics and Other, which overlap heavily with the seed words and with high-frequency compensation terms, so a much shorter seed-based dictionary would likely reproduce the same trends. The authors should benchmark the lexicon against human-coded proxy-statement or conference-call text, and compare precision/recall with the seed-word list alone and with Zhang (2022) or other existing keyword lists.
  3. [Section III (Word2Vec; Final Word lists)] The paper does not report key word2vec hyperparameters (e.g., window size, minimum count, number of epochs, subsampling) and does not provide a sensitivity analysis for the cosine-similarity threshold of 0.5. Since the threshold determines which 7,018 terms enter the screening stage and which are ultimately retained, the authors should report the full hyperparameter configuration and show that the final lexicon is robust to reasonable threshold choices (e.g., 0.4-0.6) and to alternative embedding specifications.
  4. [Section IV (An Application Using Our Textual Data)] The BERT evaluation on 10-Q sentences selects the probability threshold (86 percent) that maximizes the F1 score on the same 1,000 manually labeled sentences used to report the performance metrics, so the reported precision, recall, and F1 are optimistic. The authors should use a separate validation set for threshold selection and a distinct test set for evaluation, or report cross-validated metrics. The manual labeling of the 1,000 sentences also lacks reliability statistics.
minor comments (5)
  1. [Table 3, Panel E] The final lexicon includes terms such as 'people' and 'personally' that appear overly broad for HC disclosure measurement in general documents; the authors should either document the face-validity rationale for retaining these terms or provide an example of how they behave in out-of-sample applications.
  2. [Table 3, Panel D] The keyword 'unpaid_leaf' appears to be a typo for 'unpaid_leave'; since the lexicon is released as a data product, this should be corrected in both the table and the machine-readable file.
  3. [References] The in-text citation 'Morrow Sodali (2019)' is not listed in the reference list, and the 'CPA Canada' reference lacks publication year and URL details; these should be completed.
  4. [Appendix] The supplemental data and code link is given only as a file name ('ISYS-2023-023_HC_Disclosures_Data_and_Code.zip'); provide a persistent URL or DOI so that the stated data-availability commitment is actionable.
  5. [Figures 2 and 3] The three-dimensional cluster plots are difficult to read in black and white; consider adding labeled legends, separate two-dimensional projections, or accessible color palettes.

Circularity Check

1 steps flagged · score 2.0 of 10

Lexicon construction is self-contained; minor threshold-fitting in the BERT illustration.

  1. fitted input called prediction [Section IV, 'An Application Using Our Textual Data' (BERT 10-Q evaluation paragraph)]
    "For this exercise, we manually label 1,000 sentences, 437 of which are confirmed HC sentences. We use the fine-tuned model to predict whether a sentence is related to HC, at a probability threshold that maximizes the F1 score. At the probability threshold of 86 percent, the corresponding precision, recall, and F1 score are 85.16 percent, 85.35 percent, and 85.26 percent, respectively, for the HC class. These metrics indicate that the fine-tuned model performs impressively well on out-of-sample classification."

    The probability threshold is a decision-rule parameter fitted to the exact 1,000 manually labeled 10-Q sentences by maximizing F1 on those labels. The precision, recall, and F1 reported immediately afterward are computed on that same set at the fitted threshold, so the reported 'out-of-sample' metrics are the best achievable on the evaluation set rather than unbiased predictions. This is a fitted input called prediction, but it is confined to the illustrative BERT application and does not affect the word2vec lexicon construction or the 10-K held-out BERT test.

full rationale

The core deliverable—the 1,285-term HC lexicon—is produced by a semi-supervised word2vec expansion from externally sourced seed words (SASB/SHRM categories and manual reading) on a corpus of 10-K HC disclosures identified by SEC mandate, not by the lexicon itself. The lexicon is therefore not defined in terms of the outcome it is later used to measure. The proxy-statement application is an out-of-sample face-validity check: the lexicon is applied to documents (1994–2022 proxies) not used in training, and the observed 2007/2010/2020 spikes correspond to independent regulatory and societal events. The BERT fine-tuning uses an 80/20 split with held-out sentences, so the 10-K test metrics are genuine out-of-sample predictions. The only circular element is in the 10-Q BERT illustration: the probability threshold is optimized on the 1,000 manually labeled 10-Q sentences, and the same sentences are then used to report precision/recall/F1, so those numbers are not an unbiased out-of-sample evaluation. This is a minor methodological leak in a secondary illustration, not a load-bearing step in the construction or primary validation of the lexicon. Self-citations to Demers, Wang, and Wu (2024) concern a companion descriptive study and are not load-bearing. Overall circularity is low.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

No new physical or theoretical entities are introduced; the lexicon, data, and code are artifacts, not invented entities in the sense of this ledger. The free parameters are modeling choices that materially shape the final lexicon, and the axioms are the assumptions on which the word2vec expansion, corpus labeling, and manual screening rely.

free parameters (3)
  • Word2Vec cosine similarity threshold = 0.5 (absolute value)
    Words with cosine similarity at least 0.5 or at most -0.5 to seed words enter the initial list; chosen by hand based on prior studies (Kee 2019; Erfani et al. 2021), not tuned to this corpus.
  • Word2Vec embedding dimension = 100
    The paper states that word2vec represents each word using a vector of 100 numbers; this hyperparameter affects similarity neighborhoods and is not justified for this corpus.
  • BERT classification probability threshold for 10-Q evaluation = 0.86
    The threshold is chosen to maximize F1 on the same 1,000 manually labeled 10-Q sentences used for evaluation, so the reported 85% F1 is in-sample for the threshold.
assumptions (5)
  • domain assumption Distributional hypothesis: words occurring in similar contexts have similar meanings
    Invoked in Section III 'Word2Vec' and attributed to Harris (1954); the entire lexicon expansion depends on this.
  • domain assumption The hand-collected HC disclosure sections from 10-Ks are unambiguously and exclusively HC-related
    Section III sample selection states that training on this corpus 'ensures that our lexicon has high construct validity.' If sections contain non-HC boilerplate, the learned contexts are contaminated.
  • domain assumption The five seed categories based on SASB, SHRM, and manual reading cover the multidimensional HCM construct
    Section III 'Initial seed words' chooses the taxonomy and seed words that determine what the algorithm can discover.
  • ad hoc to paper Words with cosine similarity at or above 0.5 to seed words are meaningfully related to HCM
    Section III 'Final Word lists' applies a global threshold uniformly, justified by citation rather than by validation on this corpus.
  • ad hoc to paper Manual screening of 7,018 candidates to 1,285 final terms accurately separates HC-related from non-HC words
    Section III 'Final Word lists' reports the manual screening step but no inter-coder reliability or second-coder audit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Measuring Corporate Human Capital Disclosures: Lexicon, Data, Code, and Research Opportunities." pith.science (2026). https://pith.science/paper/JEQ6A7X5

@misc{pith2026250610155,
  author       = {Pith},
  title        = {Pith review of: Measuring Corporate Human Capital Disclosures: Lexicon, Data, Code, and Research Opportunities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEQ6A7X5}},
  note         = {Machine review of arXiv:2506.10155}
}
read the original abstract

Human capital (HC) is increasingly important to corporate value creation. Unlike other assets, however, HC is not currently subject to well-defined measurement or disclosure rules. We use a machine learning algorithm (word2vec) trained on a confirmed set of HC disclosures to develop a comprehensive list of HC-related keywords classified into five subcategories (DEI; health and safety; labor relations and culture; compensation and benefits; and demographics and other) that capture the multidimensional nature of HC management. We share our lexicon, corporate HC disclosures, and the Python code used to develop the lexicon, and we provide detailed examples of using our data and code, including for fine-tuning a BERT model. Researchers can use our HC lexicon (or modify the code to capture another construct of interest) with their samples of corporate communications to address pertinent HC questions. We close with a discussion of future research opportunities related to HC management and disclosure.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 34 canonical work pages

  1. [9]

    Journal of Financial Economics 101 (3): 621–640

    Does the stock market fully value intangibles? Employee satisfaction and equity prices. Journal of Financial Economics 101 (3): 621–640. doi:10.1016/j.jfineco.2011.03.021. Ehnert, I., S. Parsa, I. Roper, M. Wagner, and M. Muller-Camen

    work page doi:10.1016/j.jfineco.2011.03.021 2011

  2. [11]

    Journal of Construction Engineering and Management 147 (12): 04021175

    An empirical analysis of risk similarity among major transportation projects using natural language processing. Journal of Construction Engineering and Management 147 (12): 04021175. doi:10.1061/(ASCE)CO.1943-7862.0002206. Geiger, R. S., D. Cope, J. Ip, M. Lotosh, A. Shah, J. Weng, and R. Tang

    work page arXiv 1943

  3. [12]

    Garbage in, garbage out

    “Garbage in, garbage out” revisited: What do machine learning application papers report about human-labeled training data? Quantitative Science Studies 2 (3): 795–827. doi:10.1162/qss_a_00144. Green, T. C., R. Huang, Q. Wen, and D. Zhou

    doi:10.1162/qss_a_00144

  4. [13]

    Journal of Financial Economics 134 (1): 236–251

    Crowdsourced employer reviews and stock returns. Journal of Financial Economics 134 (1): 236–251. doi:10.1016/j.jfineco.2019.03.012. GRI

    work page doi:10.1016/j.jfineco.2019.03.012 2019

  5. [15]

    The Quarterly Journal of Economics 134 (4): 2135–2202

    Firm-level political risk: Measurement and effects. The Quarterly Journal of Economics 134 (4): 2135–2202. doi:10.1093/qje/qjz021. Henry, E

    work page doi:10.1093/qje/qjz021

  6. [18]

    Review of Accounting Studies 21 (4): 1005–1045

    The benefits of specific risk-factor disclosures. Review of Accounting Studies 21 (4): 1005–1045. doi:10.1007/s11142-016-9371-1. Huang, A. H., H. Wang, and Y. Yang

    work page doi:10.1007/s11142-016-9371-1

  7. [19]

    Contemporary Accounting Research 40 (2): 806–841

    FinBERT: A large language model for extracting information from financial text. Contemporary Accounting Research 40 (2): 806–841. doi:10.1111/1911-3846.12832. Huang, X., S. H. Teoh, and Y. Zhang

    arXiv 1911

  8. [21]

    ISO 30414:2018 Human resource management - Guidelines for internal and external human capital reporting. ISO. https://www.iso.org/standard/69338.html. Kee, T

    work page 2018

Show all 41 references
  1. [22]

    In 2019 IEEE International Conference on Big Data (Big Data), 5536–5543

    Peer firm identification using word embeddings. In 2019 IEEE International Conference on Big Data (Big Data), 5536–5543. doi:10.1109/BigData47090.2019.9006438. Li, F

    2019

  2. [23]

    Journal of Accounting and Economics 45 (2)

    Annual report readability, current earnings, and earnings persistence. Journal of Accounting and Economics 45 (2). Economic Consequences of Alternative Accounting Standards and Regulation: 221–247. doi:10.1016/j.jacceco.2008.02.003. Li, K., F. Mai, R. Shen, and X. Yan

    2008 doi

  3. [24]

    The Review of Financial Studies 34 (7): 3265–3315

    Measuring corporate culture using machine learning. The Review of Financial Studies 34 (7): 3265–3315. doi:10.1093/rfs/hhaa079. Liang, C., B. Lourie, A. Nekrasov, and I. S. Yoo

    doi

  4. [25]

    The Accounting Review 86 (4): 1383–1414

    What makes conference calls useful? The information content of managers’ presentations and analysts’ discussion sessions. The Accounting Review 86 (4): 1383–1414. https://www.jstor.org/stable/23045606. Mayew, W. J., and Y. Zhang

  5. [26]

    Accounting Forum 42 (1): 130–141

    A study of human capital reporting in the United Kingdom. Accounting Forum 42 (1): 130–141. doi:10.1016/j.accfor.2017.11.001. Michaelides, A., and N. Vafeas

    2017 doi

  6. [27]

    doi:10.1016/j.jaccpubpol.2023.107083

    Chief Human Resource Officers and accounting disclosures: Illuminating the firm’s most important asset or window dressing? Journal of Accounting and Public Policy 42 (3): 107083. doi:10.1016/j.jaccpubpol.2023.107083. 29 Mikolov, T., K. Chen, G. Corrado, and J. Dean

    2023

  7. [29]

    Applied Economics Letters 10 (9): 549–552

    Effects of employee training on the performance of North-American firms. Applied Economics Letters 10 (9): 549–552. doi:10.1080/1350485032000100297. Morrow Sodali

    doi

  8. [30]

    https://morrowsodali.com/insights/institutional- investor-survey-2019

    Institutional investor survey. https://morrowsodali.com/insights/institutional- investor-survey-2019. Muslu, V., S. Radhakrishnan, K. R. Subramanyam, and D. Lim

    2019

  9. [32]

    https://cooleypubco.com/2022/06/15/petition-financial-disclosure-human-capital/

    Working group petitions SEC to mandate financial disclosure requirements for human capital. https://cooleypubco.com/2022/06/15/petition-financial-disclosure-human-capital/. Rayton, B. A

    2022

  10. [33]

    Journal of Corporate Finance 9 (3): 333–352

    Firm performance and compensation structure: performance elasticities of average employee compensation. Journal of Corporate Finance 9 (3): 333–352. doi:10.1016/S0929- 1199(02)00017-2. SASB

    doi

  11. [35]

    The Journal of Finance 78 (3): 1449–1498

    Firm-level climate change exposure. The Journal of Finance 78 (3): 1449–1498. doi:10.1111/jofi.13219. SEC

    doi

  12. [37]

    doi:10.1007/s10551-019-04289-3

    Human resource disclosures in UK corporate annual reports: To what extent do these reflect organisational priorities towards labour? Journal of Business Ethics 169 (3): 475–497. doi:10.1007/s10551-019-04289-3. Wang, Y.-Y., and T. Wang

    doi

  13. [38]

    Journal of Information Systems 36 (3): 211–217

    Python code and illustrative crisis management data from Twitter. Journal of Information Systems 36 (3): 211–217. doi:10.2308/ISYS-2022-011. Webersinke, N., M. Kraus, J. A. Bingler, and M. Leippold

    2022 doi

  14. [39]

    ClimateBert: A pretrained language model for climate-related text. arXiv. doi:10.48550/arXiv.2110.12010. Zhang, M

    doi arXiv

  15. [40]

    SSRN Scholarly Paper

    Determinants and consequences of human capital management disclosure. SSRN Scholarly Paper. Rochester, NY: Social Science Research Network. https://papers.ssrn.com/abstract=3961202. 30 FIGURE 1 Distribution of Cosine Similarity Scores This graph plots the distribution of the c...

    2020

  16. [41]

    35 FIGURE 5 Trend of HC Disclosures in Proxy Statements over 1994-2022 The left vertical axis shows the absolute HC keyword count

    On the horizontal axis, “1” represents the first 10 days of this one-year period. 35 FIGURE 5 Trend of HC Disclosures in Proxy Statements over 1994-2022 The left vertical axis shows the absolute HC keyword count . The right vertical axis tracks the relative HC keyword count as...

    1994

  17. [1113]

    doi:10.2308/accr-50684. ISO

    doi

  18. [1954]

    WORD 10 (2–3): 146–162

    Distributional structure. WORD 10 (2–3): 146–162. doi:10.1080/00437956.1954.11659520. Haslag, P. H., B. A. Sensoy, and J. T. White

    1954

  19. [2003]

    Journal of Human Resource Costing & Accounting 7 (1): 9–20

    Intellectual capital disclosure in Canadian corporations. Journal of Human Resource Costing & Accounting 7 (1): 9–20. doi:10.1108/eb029076. Bourveau, T., M. Chowdhury, A. Le, and E. Rouen

    doi

  20. [2007]

    Australian Accounting Review 17 (42): 71–83

    Intellectual capital reporting practices of the top australian firms. Australian Accounting Review 17 (42): 71–83. doi:10.1111/j.1835-2561.2007.tb00445.x. Vithana, K., T. Soobaroyen, and C. G. Ntim

    2007

  21. [2008]

    doi:10.1177/0021943608319388

    Are investors influenced by how earnings press releases are written? The Journal of Business Communication (1973) 45 (4): 363–407. doi:10.1177/0021943608319388. Hoitash, R., and U. Hoitash

    1973 doi

  22. [2009]

    Canadian Journal of Administrative Sciences 26 (1): 71–88

    Attributes of social and human capital disclosure and information asymmetry between managers and investors. Canadian Journal of Administrative Sciences 26 (1): 71–88. doi:10.1002/cjas.89. CPA Canada. The rise of the social pillar: The ‘S’ in ESG. https://www.cpacanada.ca/en/bu...

    doi

  23. [2010]

    Journal of Human Resource Costing & Accounting 14 (4): 262–285

    Human capital, value creation and disclosure. Journal of Human Resource Costing & Accounting 14 (4): 262–285. doi:10.1108/14013381011105957. Bernstein, A., and L. Beeferman

    doi

  24. [2011]

    Journal of Accounting Research 49 (2): 309–346

    Large-sample evidence on firms’ year-over-year MD&A modifications. Journal of Accounting Research 49 (2): 309–346. doi:10.1111/j.1475- 679X.2010.00396.x. Chen, J., E. Demers, and B. Lev

    2010 arXiv

  25. [2013]

    arXiv:1301.3781 [cs]

    Efficient estimation of word representations in vector space. arXiv:1301.3781 [cs]. http://arxiv.org/abs/1301.3781. Miller, S

    arXiv

  26. [2014]

    Management Science 61 (5): 931–948

    Forward-looking MD&A disclosures and the information environment. Management Science 61 (5): 931–948. doi:10.1287/mnsc.2014.1921. Nichite, P

    2014

  27. [2016]

    28 The International Journal of Human Resource Management 27 (1): 88–108

    Reporting on sustainability and HRM: A comparative study of sustainability reporting practices by the world’s largest companies. 28 The International Journal of Human Resource Management 27 (1): 88–108. doi:10.1080/09585192.2015.1024157. Erfani, A., Q. Cui, and I. Cavanaugh

  28. [2018]

    Management Science 64 (12): 5899–5924

    Oh what a beautiful morning! Diurnal influences on executives and analysts: Evidence from conference calls. Management Science 64 (12): 5899–5924. doi:10.1287/mnsc.2017.2888. Cormier, D., W. Aerts, M.-J. Ledoux, and M. Magnan

    2017

  29. [2019]

    BERT: Pre-training of deep bidirectional transformers for language understanding. arXiv. doi:10.48550/arXiv.1810.04805. Drei, A., T. Le Guenedal, F. Lepetit, V. Mortier, T. Roncalli, and T. Sekine

    doi arXiv

  30. [2020]

    https://sasb.org/wp- content/uploads/2020/12/Human-Capital_Preliminary-Framework_2020-December_FINAL.pdf

    Preliminary Framework on Human Capital and the SASB Standards. https://sasb.org/wp- content/uploads/2020/12/Human-Capital_Preliminary-Framework_2020-December_FINAL.pdf. Sautner, Z., L. Van Lent, G. Vilkov, and R. Zhang

    2020

  31. [2021]

    Review of Finance: rfab036

    Language and domain specificity: A Chinese financial sentiment dictionary. Review of Finance: rfab036. doi:10.1093/rof/rfab036. Edmans, A

    doi

  32. [2022]

    Journal of Information Systems 36 (2): 161–172

    A measure of firm complexity: Data and code. Journal of Information Systems 36 (2): 161–172. doi:10.2308/ISYS-2021-041. Hope, O.-K., D. Hu, and H. Lu

    2021 doi

  33. [2023]

    Journal of Information Systems 37 (1): 169–188

    Toward automating shredding nonprofit XML files: The case of IRS Form 990 data. Journal of Information Systems 37 (1): 169–188. doi:10.2308/ISYS-2022-

    2022 doi

Pith tools

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