Skip to content

AI uses for records

Table footnote reasoning dataset: linking each table value to the note that governs it

By SourceX Editorial · Updated

10 sections

Short answer

Link them explicitly. Each item should cite the table cell by page, region and header path, the footnote marker and its scope (cell, row, column, table or remote), the footnote region, and the relation the note creates, such as restating, narrowing scope, changing scale or redefining the metric. Ask questions that cannot be answered from the cell alone, add control items where the note is irrelevant, and grade evidence as well as answers.

Value drivers

Table footnote reasoning

This page covers table footnote reasoning dataset: linking each table value to the note that governs it.Illustrative

Key takeaways

  • A footnote often changes what a table value means, so the answer and the cited evidence both depend on reading it.
  • Store the cell, the marker and its scope, the footnote region and a typed relation as separate fields in every record.
  • Born-digital sources such as XBRL footnote links and HTML header markup give silver labels to check against the rendered page.
  • Write questions for every relation type, plus controls where the note does not apply and items where it is missing.
  • Split by issuer and template and score answer, link, relation and evidence separately.

Why can't a model read the number without its footnote?

Because the footnote often changes what the number means. It can narrow the scope of a figure, change its unit or scale, mark it as restated or preliminary, redefine the metric, or point to a note elsewhere that adjusts it. A model that reads the cell alone answers a different question from the one the report is answering.

Public table question answering sets show the size of the problem without isolating it. TAT-QA pairs 16,552 questions with 2,757 hybrid contexts of tables and text taken from real financial reports. FinQA records hold the text before the table, the table, the text after it, the question and a reasoning program. Both make models read text around tables, but neither record format, as published, has a field that says which marker on which cell points to which note.

That missing link is the gap this specification fills. Each item names the cell, the marker, the footnote region and the relation between them, so a grader can tell whether a model got the right answer for the right reason, and a trainer can teach the link rather than hoping it emerges.

Where do footnotes sit, and how do markers attach to cells?

A marker can attach to one cell, a whole row through its row header, a whole column through its column header, the table title, or the table as a whole, and the scope of the note follows the marker. The footnote text itself can sit directly under the table, at the bottom of the page, at the end of a section, in a numbered note such as Note 7, or in a separate file.

At a glance

Where do footnotes sit, and how do markers attach to cells

Layout datasets locate footnotes but do not link them. DocLayNet provides bounding-box labels for 11 layout classes, including Footnote and Table, on 80,863 human-annotated pages. Its labeling guide describes a footnote as typically small text at the bottom of a page with a number or symbol referred to in the text above, labels several footnotes on one page as separate clusters, and tells annotators that a page footer is not a footnote and that a table caption is not part of the table cluster.

Those rules are a good starting point for region labels. The reasoning dataset then adds what layout labels leave out: the marker's position inside the table grid, the scope it applies to, and a typed edge from the marker to the specific footnote region.

  • Cell scope: a dagger on one value, such as a single quarter that was restated.
  • Row scope: a marker on a row header, such as a product line that excludes one region.
  • Column scope: a marker on a column header, such as a prior-year column on a revised basis.
  • Table scope: a marker on the title or a note directly under the table that applies to every value.
  • Remote scope: a cross-reference such as see Note 7, where the governing text is pages or files away.

Born-digital filings and web tables often carry the link in markup, which makes them a cheap source of silver labels to check against the rendered page. Scanned documents carry none, and the link has to be drawn by annotators.

XBRL 2.1, the standard behind many structured financial filings, defines footnote links in which arcs connect fact locators to footnote resources, usually with the fact-footnote arcrole. In a filing that uses them, a reported value and its footnote are joined in the data itself. The SEC publishes Financial Statement and Notes data sets that extract the text and detailed numeric information in financial statements and their notes from XBRL exhibits to corporate filings.

For HTML tables, the WHATWG HTML Standard defines how data cells are associated with header cells, either through an explicit headers attribute or through an algorithm that scans the row and column using the scope attribute. That gives a principled way to record the row and column header path for each cell, which the specification needs in order to decide whether a row or column marker covers it.

Treat all of these as inputs to verify, not as ground truth. Filers can attach notes inconsistently, HTML tables are often built for appearance rather than structure, and the rendered page is what a human reader, and a multimodal model, actually sees.

What should the table and footnote reasoning dataset specification include?

Each record ties one question to the minimum set of page regions needed to answer it, with the cell, marker, footnote and relation stored as separate fields. The table below is the field list; use it as a delivery schema in a data request or as the annotation form for an internal labeling job.

Two design choices matter most. Store the value as printed and as normalized, because many footnotes change scale or units. And store the relation type as a closed list, so graders and analysts can break results out by the kind of reasoning the note demands.

What should the table and footnote reasoning dataset specification include?
FieldWhat it holdsCheck before acceptance
source_file and versionDocument id, file hash, version or amendment number and issue dateHash matches the delivered file; amendments are separate records
page and table regionPage number, table bounding box, caption region and table idBox encloses the grid only; caption stored apart, as layout guides require
cell addressRow and column index, row header path and column header pathHeader paths agree with the rendered table, not only the markup
valueText as printed, normalized number, unit, scale and periodScale words such as in thousands are applied once, not twice
markerSymbol or number, its location in the grid and its scope: cell, row, column, table or remoteScope is set by where the marker sits, confirmed by a second annotator
footnoteFootnote id, region box, full text and OCR confidenceLow-confidence text is flagged, not silently used as gold
relationOne of restates, restricts scope, changes unit or scale, redefines, flags estimate, refers elsewhere, no effectRelation chosen from the closed list with a one-line rationale
cross-file linkPointer to a note, schedule or exhibit in another file, with its own page and regionLinked file is in the delivery and the pointer resolves
question and answerQuestion text, gold answer, unit and the arithmetic or rule usedAnswer changes if the footnote is ignored, unless the item is a control
evidence set and reviewEvery region required for the answer, the reviewer role and the review dateNo answer depends on a region missing from the evidence set

Which footnote relations should the questions test?

Write questions for each relation type, because each one breaks a different shortcut. A set where every footnote is a restatement teaches a model to always prefer revised figures, which is wrong for rounding notes and for questions about what was originally reported.

The examples below are illustrative and use counts, durations and percentages only. Each row shows the wrong answer a model gives when it reads the cell without the note, which is the failure the item is designed to catch.

Which footnote relations should the questions test?
RelationWhat the note doesIllustrative questionAnswer if the note is ignored
RestatesSays a prior-period figure was revised on a new basisHow did units shipped change year over year on a comparable basis?Compares the revised year with an unrevised one
Restricts scopeExcludes a segment, region or class of items from the totalDoes the headcount figure include contractors?Assumes the total covers all workers
Changes unit or scaleStates that figures are in thousands or in a different unitHow many service tickets were closed in the year?Off by the scale factor
RedefinesGives a company-specific definition of the metricWhat counts as an active customer in this table?Uses a generic definition
Flags estimateMarks a value as preliminary, unaudited or projectedIs the fourth-quarter return rate final?Treats the figure as final
Refers elsewherePoints to a numbered note that adjusts the valueWhat share of backlog is expected to ship within 12 months?Answers from the table and misses the note
No effectRounding or presentation note with no bearing on the questionWhat was the median resolution time in days?Control item: the right answer ignores the note

How should evidence regions be recorded when the governing note sits on another page or file?

The evidence is the smallest set of regions a careful reader needs: usually the cell and its footnote, sometimes the row or column header that carries the marker, and sometimes a note on another page or in another file. Record every one of them with page, box and file, and grade whether the model cited them, not only whether its number matched.

Images and attachments add two complications. In scanned reports the marker can be a faint superscript that optical character recognition drops, so keep the OCR confidence for both the marker and the footnote text and route low-confidence items to a human check. And when the governing note lives in an exhibit or schedule, the dataset must deliver that file too; a pointer to a document the model cannot see turns a reasoning item into a guessing item.

Order matters when there are several notes. Record the evidence set as an ordered chain, for example cell, column header marker, footnote, then the numbered note it refers to, so a grader can see where a model's reasoning broke rather than only that it did.

Where a born-digital source exists, align the rendered page with the structured link, such as an XBRL footnote arc, and keep both. Disagreements between the two are worth keeping as hard cases, labeled as such.

Illustrative worked example: a shipments table with a restated column

This example is illustrative; the company, counts and notes are invented to show the method, and no results are reported. A two-column operations table lists units shipped by product line for the current and prior year, with values in thousands stated in the caption. The prior-year column header carries a dagger.

The dagger note below the table says the prior year was restated to move one product line into a different segment. A second note, marked with an asterisk on the accessories row, says accessory units exclude returns processed after the year end. A third note says totals may not sum because of rounding.

The item asks: on a comparable basis, by what percentage did units shipped in the main product line grow? The evidence set is the two cells, the dagger on the column header, the dagger footnote region and the caption that sets the scale. The relation is restates. A control item on the same table asks for the median, which the rounding note does not affect, with relation no effect.

A model that cites the cell and the restatement note and uses the restated prior-year figure passes both answer and evidence checks. A model that reaches the right percentage while citing only the cells passes the answer and fails attribution, which is the signal that it may have got lucky.

When does a nearby footnote not govern the value?

A nearby note does not govern a value when its marker is not on that value's cell, row, column or table, when it is a page footer rather than a footnote, or when it is a presentation note with no effect on the question. Items that test these cases stop a model from applying every note it sees.

Reused symbols are the classic trap. Two tables on one page can each use an asterisk with a different meaning, and the right note is the one under the table that holds the marker. Notes that apply only to a cell should not be stretched to its whole row, and remote references should be followed only when the marker points to them.

Include a few items where the governing note is missing or illegible and the expected behavior is to say so. A model that invents a plausible restatement is more dangerous than one that answers from the cell and flags the gap.

How should a footnote reasoning set be scored, split and bounded?

Score four things separately: the answer, the marker-to-note link, the relation type and the evidence set, plus correct abstention on items with missing notes. Report each by relation type and by document source type, such as born-digital, scanned or HTML, so a buyer can see where the model fails.

Grade evidence with region overlap, not exact boxes. Two careful annotators rarely draw identical rectangles, so count a cited region as correct when it overlaps the gold region above a threshold you set and publish, and check that the cited region holds the marker or note text the item names.

Split by issuer and document template, not by item. Reports from one company reuse wording, layouts and note numbering across years, and a random split lets a model learn the template rather than the reasoning. DocLayNet makes the same point for layouts by fixing train, test and validation sets to avoid leakage of unique layout styles.

Know the limits. The specification covers single tables with notes and simple cross-references; it does not cover multi-table reconciliation, notes written in a language other than the table's, or charts. If your use case needs those, extend the relation list and the evidence set rather than forcing them into these fields.

Where can real reports with tables and governing notes come from?

Public filings are useful for building and testing the annotation process, but private operational reports, such as monthly operations packs, vendor scorecards and board reporting tables, carry the footnote habits that a business deployment will meet. They also come with versions and amendments that make restatement items real rather than invented.

Report sets like this are never stocked: SourceX locates them at U.S. businesses for each request, and the buyer describes the scope, volume, language, format, timing and licensing terms needed, for instance reports with tables, notes and their exhibits delivered together. Delivery waits until the company that owns the reports has approved what is in scope, agreed the terms and signed.

Frequently asked questions

What is a table footnote reasoning dataset?

It is a set of questions over real documents where the correct answer depends on a footnote that governs a table value. Each item records the cell, the footnote marker and its scope, the footnote text region and the relation between them, so models can be graded on reasoning and evidence rather than on the final number alone.

How is this different from table structure recognition data?

Table structure recognition labels the grid: cells, rows, columns and spanning headers. A footnote reasoning set assumes the grid is known and adds the link from a marker to the note that changes how a value is read, plus questions that cannot be answered from the cell alone.

Can XBRL footnote links be used as ground truth?

They are a strong starting point when a filing uses them, because XBRL footnote links join facts to footnote resources in the data. Check them against the rendered page, since filers apply notes inconsistently, and keep disagreements as labeled hard cases.

How should scanned documents with faint markers be handled?

Keep the OCR confidence for both the marker and the footnote text, and send low-confidence items to a human reviewer before they become gold. If the marker cannot be read reliably, either drop the item or label it as one where the correct behavior is to flag the uncertainty.

Should every item include a footnote that matters?

No. Include control items where a nearby note, such as a rounding note, does not affect the answer, and items where the note is missing. Without them a model learns to apply every note it sees, which fails on real documents.

Sources

  • TAT-QA contains 16,552 questions associated with 2,757 hybrid contexts of tables and text from real-world financial reports. Source
  • FinQA records hold the text before the table, the table, the text after the table, the question and a reasoning program. Source
  • DocLayNet provides bounding-box layout labels for 11 classes on 80,863 human-annotated pages, with fixed train, test and validation sets to avoid leakage of unique layout styles. Source
  • The DocLayNet labeling guide defines a footnote as typically small text at the bottom of a page with a number or symbol referred to above, labels several footnotes as separate clusters, and excludes page footers and captions. Source
  • XBRL 2.1 defines footnote links in which arcs connect fact locators to footnote resources. Source
  • The SEC's Financial Statement and Notes data sets provide text and detailed numeric information from financial statements and their notes, extracted from XBRL exhibits. Source
  • The WHATWG HTML Standard defines how data cells are assigned header cells, through the headers attribute or a scanning algorithm using the scope attribute. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies