Disclosure Mining Techniques: A Forensic Methodology

Disclosure mining is the systematic extraction of evidence from a company's public record (SEC filings, earnings call transcripts, proxy statements, press releases) so that what management says can be tested against what the company demonstrably does. The techniques fall into three families: lexical methods that count and classify words, statistical and machine-learning methods that model context, and structured-data verification that anchors narrative claims to mandatory, audit-traceable tables. The most reliable practice combines all three, with the third acting as the arbiter.

Lacunaindex

Test Corporate Claims Against Delivery

Lacuna Index analyzes public disclosures to assess how companies deliver on their claims, using evidence from filings, calls, and public statements.

Explore Lacuna Index

TL;DR

Dictionary-based word counts are efficient at scale but misclassify context-dependent language, so hybrid methods that add statistical content analysis are now preferred in the literature.
Narrative tone is not static: earnings calls and SEC filings can carry meaningfully different messages from the same company.
Mandatory disclosure tables, such as Exhibit 95 for mine safety, provide a baseline of objectivity that sentiment scores cannot supply.
A defensible workflow runs in five stages: collect, parse, extract claims, ground claims against hard data, and log every scoring decision.
The search results for this topic are dominated by mining-industry regulation; the analytical discipline of mining disclosures is a separate subject, covered here.

1. Two meanings of the phrase, and why the distinction matters
2. The evolution of techniques
3. Why mandatory tables are the anchor
4. A five-stage workflow for disclosure mining
5. Comparing technique families
6. Common failure modes
7. Perspective: what a delivery-based reading requires
8. Applying the techniques: a worked structure
9. Where Lacuna Index fits
FAQ
Sources

1. Two meanings of the phrase, and why the distinction matters

Search results for this term are populated largely by guidance on disclosure by mining companies: the SEC's property disclosure rules, Canada's National Instrument 43-101, and mine-closure transparency studies. That material is relevant to one part of this article, because mining registrants produce some of the most rigid disclosures in U.S. securities law. The broader subject, however, is disclosure mining as an analytical technique: the use of text and data extraction to measure the distance between corporate narrative and corporate delivery.

The two meanings meet in a useful way. Mining-sector disclosures illustrate what a disclosure looks like when management discretion is removed, and that contrast is the foundation of a forensic reading of every other disclosure.

2. The evolution of techniques

2.1 Lexical methods: dictionaries and word counts

The first generation of textual analysis in finance relied on word lists. A filing is tokenized, each word is matched against a dictionary of positive, negative, uncertain or litigious terms, and the proportions become a tone score. The survey of the textual analysis literature attributes the growth of this research to two developments: large volumes of unstructured text becoming electronically available, and the accessibility of filing archives such as EDGAR.

Filing words classified into a tone score

The approach has clear strengths: it is transparent, reproducible and cheap to run across thousands of filings. Its weakness is context. The same survey notes that general-purpose dictionaries (the General Inquirer is the example given) cannot distinguish the meaning of a word across settings: "increase" is favorable beside sales and unfavorable beside costs. A score built on such counts will misclassify precisely the sentences that matter most in a forensic review, those containing forward-looking judgment.

2.2 Statistical and hybrid methods

The second generation adds statistical content analysis, in which the model learns which word patterns accompany particular outcomes instead of assuming a fixed polarity per word. Reviews of the field indicate that dictionary approaches remain efficient for large datasets but can miss nuance that statistical content mining identifies. Hybrid designs combine dictionary-based thematic tagging (to identify the topic of a sentence: guidance, capital allocation, cost reduction) with statistical pattern analysis (to resolve what the sentence is asserting about that topic).

A broader survey of text as data in finance and accounting describes how these tools have become increasingly popular for extracting meaning from unstructured text. For practitioners, the practical lesson is that method selection should follow the question: counting is adequate for screening, while claim-level assessment demands context-aware models.

The evolution of retail-facing AI tools follows the same arc. A review of what works and what to verify when using ChatGPT for stock research addresses the same underlying tension: language models extend what text analysis can read, but outputs still require verification against primary sources.

2.2 Statistical and hybrid methods — overview diagram

2.3 Large language models and the verification requirement

Large language models extend extraction to long, irregular documents, but a paper on text analysis in financial disclosures frames the central problem: prevailing methods lean on quantitative ratios that suffer from window dressing and past-looking bias, while text offers forward-looking signal that is harder to read reliably. The limitation of any statistical or neural model is that it produces plausible readings without any guarantee of correctness. The remedy is structured verification: comparing a narrative claim in an executive statement against a hard table in a filing exhibit.

3. Why mandatory tables are the anchor

Narrative sections of a filing are shaped by management judgment about what is material and how to phrase it. Mandatory tables are not. The analytical literature therefore treats them as a baseline against which narrative can be calibrated, an argument made explicitly in the financial disclosure text analysis paper cited above.

Mining registrants offer the clearest U.S. example, in two layers.

3.1 Mine safety disclosures under Dodd-Frank Section 1503

Section 1503 of the Dodd-Frank Act requires operators of coal or other mines, or issuers with a subsidiary that operates one, to report specified mine safety information. Crowell's summary of the requirement describes the categories reported in periodic filings, including significant and substantial citations, orders and imminent danger orders, along with the total dollar value of proposed MSHA assessments and mining-related fatalities. Crowell's analysis states that no materiality threshold applies to these occurrences. A representative Exhibit 95 filed on EDGAR shows the format: the statutory preamble followed by mine-by-mine tabulations.

The forensic value lies in the removal of filtering. An ESG commentary on mine safety disclosures describes them as a teaching device for how forced transparency exposes operational risk, since the figures derive from regulator-issued violations rather than management characterization. A reviewer can read a CEO's statement that safety is a core value and then test it against the citation counts for the same period.

3.2 Technical report summaries under Regulation S-K Subpart 1300

The second layer concerns property disclosure. The SEC adopted modernized mining property disclosure rules in 2018, replacing the earlier Industry Guide 7 framework. The regulation at 17 CFR Subpart 229.1300 applies when a registrant's mining operations are material to its business or financial condition. Under Item 1302, disclosure of exploration results, mineral resources or mineral reserves must be based on information and documentation prepared by a qualified person. A statement about reserves therefore carries a named professional's attestation, which narrative language about "growth potential" does not.

The general principle extends beyond mining. Wherever a disclosure carries an attestation, a statutory format or an external issuing authority, it can serve as a grounding reference. Examples across sectors include XBRL-tagged financial statement values, auditor reports, executive compensation tables in proxy statements and segment disclosures.

4. A five-stage workflow for disclosure mining

The following sequence reflects the practice described in the literature and is the structure on which an auditable process depends.

Collection. Retrieve filings by company identifier, form type and date range from EDGAR, then add transcripts, press releases and proxy materials. Open-source connectors exist for this purpose, such as the documented EDGAR search and read tools, which search filings by ticker or CIK and fetch full text by accession number.
Parsing. Segment each document into sections (business description, risk factors, MD&A, exhibits, prepared remarks, question-and-answer) so that results remain attributable to a location.
Claim extraction. Identify sentences that assert a commitment, target, causal explanation or performance outcome. This is the stage at which hybrid models outperform dictionaries.
Grounding. Match each claim to a hard reference: reported figures, mandatory exhibit tables, subsequent filings or regulator data. Claims without a possible grounding are recorded as unverifiable rather than discarded.
Logging. Record the source passage, the grounding evidence and the scoring rationale for every assessment, so a third party can reproduce the result.

4.1 Cross-source comparison

A single document rarely reveals the gap. Research on the value relevance of corporate tone, published in the Journal of Financial Reporting and Accounting, indicates that the relevance of tone varies with the information source and the methodology used. In practical terms, a company's earnings call and its SEC filing can present materially different emphasis, and the discrepancy is itself a data point. For method detail on this comparison, see the earnings call narrative analysis guide and the tone shift analysis guide.

5. Comparing technique families

Technique	Primary strength	Primary failure mode	Best use
Dictionary word counts	Transparent, cheap, reproducible	Misreads context-dependent terms	Screening large universes
Statistical content analysis	Learns context-specific patterns	Requires labeled data; opaque weights	Topic and tone modeling
Hybrid (thematic tags plus statistics)	Corrects polarity by topic	Greater design complexity	Claim-level extraction
Language-model extraction	Handles long, irregular text	Plausible but unverified output	Drafting, summarizing, candidate claims
Structured-data grounding	Objective, audit-traceable	Limited to what is mandated	Arbitrating narrative claims

No single row suffices. The pairing that produces defensible output is a context-aware extraction method joined to structured-data grounding.

6. Common failure modes

Treating tone as performance. A positive tone score measures language, not delivery. The analysis of sentiment versus performance in SEC filings sets out how to test one against the other.
Ignoring silence. Absent topics matter. A target announced in one year and omitted the next leaves no sentiment to score, yet constitutes a significant observation. The disclosure versus returns rubric addresses silent misses.
Unattributable scores. If a score cannot be traced to a passage and an evidence item, it cannot be audited.
Inconsistent treatment across issuers. Changing the rubric by company destroys comparability. Standardization is a precondition for sector benchmarks.
Over-reliance on model output. Language models require the structured verification described in section 2.3.

7. Perspective: what a delivery-based reading requires

Our work begins from a narrow premise: public records are sufficient to measure the gap between what a company claims and what it delivers, provided the process is applied identically to every issuer and every scoring decision is logged. We use no insider access and no proprietary data, which means each finding can be checked by any reader holding the same filings.

Three conclusions follow from that design. First, sentiment alone is an inadequate proxy for accountability, since tone describes how management wishes to be read. Second, the value of a disclosure to a forensic reader rises with the rigidity of its format, which is why mandatory exhibits and attested figures serve as grounding references. Third, an unasked question is as informative as an engineered answer, so the method records what remains unaddressed alongside what is stated.

The broader framework is described in our methodology walkthrough, and the logic of testing claims against records appears in the guide on how to validate corporate claims in public disclosures.

— Glen

8. Applying the techniques: a worked structure

Consider a hypothetical operator that states in an earnings call that safety performance is improving. The following steps illustrate, without reference to any actual company, how the workflow resolves the claim.

The claim extraction stage flags the sentence and its location in the prepared remarks.
The grounding stage retrieves the Exhibit 95 tabulations for the current and prior periods and compares citation and order counts.
The cross-source stage checks whether the 10-K narrative repeats, qualifies or omits the claim.
The logging stage records the passage, the table values and the conclusion (consistent, inconsistent or unverifiable).

Outside mining, the same sequence applies with different anchors: a cost-reduction target is tested against reported expense lines, a capital allocation commitment against subsequent cash flow statements, and a governance assurance against the proxy statement. Further reading on traceable extraction is available in the guide to SEC filing textual analysis with XBRL grounding, and on the interpretive side in the guide to annual report discrepancies in equity research.

9. Where Lacuna Index fits

Lacuna Index applies this methodology across covered public companies, quantifying the gap between narrative claims and delivery from public records alone. Sector benchmarks are available without a subscription, for example for Industrials, Financials and Health Care. Company-level forensic reports, with execution and narrative scores and explicit governance logs, are available through subscription. Plan details, including Cohort Pulse and Sector Sweep, are on the pricing page.

FAQ

What is disclosure mining?

Disclosure mining is the structured extraction of information from corporate disclosures such as filings, transcripts and proxy statements. In forensic use, the aim is to compare claims with verifiable delivery rather than to score tone alone.

Are dictionary-based methods still useful?

Yes, for screening large universes cheaply and reproducibly. Their limitation is context: the literature notes that general dictionaries cannot tell whether a word such as "increase" is favorable or unfavorable, so hybrid methods are preferred for claim-level work.

Why are mine safety disclosures used as a forensic example?

Because the figures are mandated and, per Crowell's summary, reported without a materiality threshold. Management cannot filter them, which makes them a clean test of narrative statements.

Do language models replace structured verification?

No. They extend the reach of extraction over long documents, but outputs still need to be checked against hard, audit-traceable tables in the filing record.