Structured GHG Disclosure Accessibility for Listed Japanese Firms: An Engineering Pilot Using EDINET and LLM-Assisted Report Extraction

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Demonstrates a methodology for auditing structured GHG tag availability in EDINET and LLM extraction accuracy; highlights schema enforcement as key research agenda.

🏢実務担当者:Shows that relying solely on voluntary reports or EDINET tags without schema enforcement leads to data that is not comparable; suggests internal controls for GHG data quality.

🏛政策担当者:Provides evidence for the need to enforce consistent schema requirements (unit, scale, concept naming) in mandatory disclosure systems like EDINET to ensure comparability.

We examine the accessibility of structured greenhouse gas (GHG) emission data for 89 major Japanese listed firms using two complementary approaches: direct extraction from EDINET statutory filings, and LLM-assisted extraction from voluntary sustainability and integrated reports. A post-extraction audit of the cached EDINET iXBRL files found a mixed structured-disclosure state: 23 of the 89 firms contained CurrentYear GHG-related `ix:nonFraction` elements with `unitRef` and `scale` attributes (including 17 with at least one Scope-specific standalone element, of which 12 contained a Scope 1 standalone element), while 19 contained only prior-year GHG-related structured elements and 47 contained no GHG-related `ix:nonFraction` elements under the inspected concept patterns. An initial version of the extraction pipeline did not consume the `unitRef` and `scale` attributes and therefore reported zero structured values; the corrected pipeline used in this version reclassifies the 89 firms into structured-current-year, prior-year-only, and no-structured-tag categories. LLM-assisted extraction was performed on voluntary PDF reports identified through prior or current acquisition for 88 of the 89 firms, producing 103 unique report-level records of which 63 contained at least one Scope value; together these covered 52 firms, with Scope 1 values extracted for 49 firms. The voluntary-report records exhibit three distinct schema-enforcement gaps on the voluntary-report side: only 11 of 49 firm-level Scope 1 records had units unambiguous from the voluntary-report record alone (14 of 49 once EDINET cross-reference cases are included); only 34 of 63 valid report-level records contained explicit page citations; and source canonicality was not verified for any record. A fourth gap surfaced as schema-shape and type-enforcement defects detected during pre-publication audit (a Scope 1 + 2 total recorded in the Scope 1 field, a unit-unknown numeric value admitted unflagged, retrieval URLs stored in a year field, and an LLM-emitted object-shaped value where a scalar was expected); a fifth gap is the consumer-side schema-awareness gap on the EDINET side described above. In the four cases where an explicit unit-labelled EDINET reference value could be matched to an LLM extraction, normalized values fell within ±10% of the reference. The pilot suggests that the binding constraints lie less in document reading itself than in unenforced schema requirements at multiple layers of the pipeline: structured tags can exist without becoming accessible comparable data unless consumer pipelines consume the attributes (`unitRef`, `scale`, `contextRef`) that make those tags meaningful, and LLM-assisted extraction can produce candidate disclosure records but does not by itself yield comparable data. We argue that comparable GHG data require schema-level enforcement on both sides of the pipeline: at the disclosure side (consistent tagging of comparability-relevant fields) and at the consumer side (parsers that consume and preserve schema attributes), together with unit normalization, evidence traceability, source accountability, schema-shape and type-enforcement, and temporal stability of the source artifacts that anchor each record. The contribution of this engineering pilot is to specify, by demonstration of where the pipeline falls short on both sides, what such an enforced infrastructure must provide.

• SSRN https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6761458first seen 2026-07-17 23:32:44