Definitions and comparisons
Enterprise acronym ambiguity dataset: what a sense label must record
By SourceX Editorial · Updated
10 sections
Short answer
Plan for collisions: in NASA's public acronym list, 6,743 of 22,278 abbreviations (30.3 percent) carry two or more meanings, and 75.9 percent of those collide inside a single source list; in the Ministry of Justice glossary, 76 of 1,151 do. An enterprise acronym ambiguity dataset must label the sense chosen, the glossary and its version, the team scope, the deciding context and whether the right move was to ask.
Document stack
Enterprise acronym ambiguity
Key takeaways
- Collisions are routine: 30.3 percent of abbreviations in NASA's list and 6.6 percent in the Ministry of Justice glossary carry more than one meaning.
- Three in four NASA collisions sit inside a single source list, so splitting documents by team narrows the problem without removing it.
- Glossaries rarely say which sense applies where: only 18 of the Ministry's 76 colliding abbreviations have a context note on every sense.
- Meanings expire: 82 Ministry entries record a rename, replacement or retirement, so every label needs the document date and the term's status.
- Score a system on picking the right sense, asking when context cannot decide, and not asking when it can.
How often does one acronym carry two meanings inside an organization?
Often enough to design for. In the acronym list NASA publishes on GitHub, 30.3 percent of distinct abbreviations have two or more meanings, and in the UK Ministry of Justice's own glossary 6.6 percent do. The first file merges eight source lists; the second is a single department's staff-maintained page, which says it is intended for internal use.
Those two counts bracket the problem. A large technical organization accumulates thousands of short forms across teams and decades, and collisions pile up; a single department curating its own list still ends up with the same letters naming a dispute process and a design document. The question for a buyer is not whether a dataset of company documents will contain ambiguous acronyms, but whether its labels record enough to grade a system that meets them.
What follows counts the collisions reproducibly, shows where they sit, and specifies the sense label and the test that follow from the counts.
Counting collisions in two public glossaries at pinned commits
We counted every abbreviation in two public lists and treated two meanings as different when their normalized wording differed. The NASA file is lists/acronyms.json in nasa/NASA-Acronyms at commit 8a1235f, the file the repository's own browser extension loads; it holds 48,789 rows from eight source lists. The Ministry list is the table in the README of ministryofjustice/acronyms at commit 3ad8771, parsed the way the repository's own publishing script parses it.
Normalization lowercases each meaning, removes text in parentheses and turns every run of other characters into a single space. A stricter variant that also drops a final s from every word longer than three characters, so plurals merge, moves the NASA share only from 30.3 to 30.0 percent. Automated counts can still mistake two wordings of one thing for two meanings, so we read samples by hand: in a seeded sample of 30 NASA collisions, 26 named at least two different things and 4 were wording variants or could not be told apart from the list alone; of all 76 Ministry collisions, 75 named different things and one paired licence recall with licence revoke.
Collision counts for both lists under the primary rule:
| Measure | NASA acronym list (8a1235f) | Ministry of Justice glossary (3ad8771) |
|---|---|---|
| Rows kept, distinct abbreviations | 48,774 rows, 22,278 abbreviations | 1,239 rows, 1,151 abbreviations |
| Abbreviations with 2 or more meanings | 6,743 (30.3%) | 76 (6.6%) |
| Abbreviations with 3 or more meanings | 3,613 | 10 |
| Most meanings for one abbreviation | 36, for CS | 4, for IA |
| Listed senses that belong to a colliding abbreviation | 64.3% | 13.2% |
| Collisions inside a single source list | 5,118 (75.9% of collisions) | Single list |
| Hand check: collisions naming different things | 26 of 30, seeded sample | 75 of 76, all read |
Same team, same letters: where the collisions sit
Most collisions are not between distant departments. In the NASA list, 5,118 of the 6,743 colliding abbreviations, 75.9 percent, carry two or more meanings inside one source list, and only 1,625 collide solely across lists. Splitting a corpus by team or system therefore narrows the problem without removing it.
The Ministry list shows how far apart the meanings of one abbreviation can be inside a single organization. ADR is alternative dispute resolution and an architectural design record. RAG is red amber green and a race advisory group, and neither is the retrieval-augmented generation an AI team would assume. JS is JavaScript and a jury summons; DR is disaster recovery and a defence request; DQ is data quality and a directions questionnaire. HMIP names two different inspectorates, of prisons and of probation, and PO is both a prison officer and a probation officer.
Engineering vocabularies collide just as readily. In NASA's list CDR includes critical design review, climate data record and command destruct receiver among its meanings, and PDR includes preliminary design review and processed data recorder. A model that resolves CDR by frequency alone will be right in a design meeting and wrong in a climate data catalog.
What the glossary itself leaves unresolved
A glossary lists senses; it rarely says which one applies where. The Ministry table has an Info column for context, and 611 of its 1,243 rows use it, yet among the 76 colliding abbreviations only 18 carry a note on every sense, 45 on some and 13 on none. For 58 of the 76, at least one sense arrives with no hint about the setting it belongs to.
Meanings also expire. A whole-word search of the Ministry's Definition and Info cells for words such as formerly, previously, replaced, no longer and superseded matched 86 rows. Read one by one, 82 record a rename, a replacement, a retired programme or a planned replacement, such as the Criminal Records Bureau replaced by the Disclosure and Barring Service, and 4 matched for other reasons. A document written before such a change uses the old sense, so a label without a date can be right and still be graded wrong.
When an ambiguous abbreviation is dangerous, organizations ban it rather than label it. The Joint Commission created its Do Not Use list in 2004 to meet a national patient safety goal requiring accredited organizations to keep a list of abbreviations not to use. Most enterprise acronyms are nowhere near that risky, which is why the workable answer for AI data is labelling: record what a reader needed to know, not only the expansion.
The sense label: what a buyer's labelling must record
Every occurrence of an acronym in an enterprise acronym ambiguity dataset needs a label that a second annotator, or a grader, can check without asking the first. Each field answers one of the counts: collisions inside one list call for a scope, unnoted senses call for the deciding evidence, and expiring meanings call for a date and a glossary version.
Fields of the sense label, with Illustrative values:
| Field | What it holds | Illustrative value | Why the grader needs it |
|---|---|---|---|
| Surface form and span | The acronym exactly as written and its position | ADR, characters 112 to 114 of a ticket comment | Ties the label to one occurrence |
| Sense inventory and version | The glossary that supplied the candidates, with a commit or date | Team glossary, revision of 2026-09-30 | Senses change; the version fixes the candidates |
| Candidate senses | Every sense the inventory lists for this form | Alternative dispute resolution; architectural design record | Shows how hard the item is |
| Chosen sense or ask | The sense that applies, or a flag that context cannot decide | Architectural design record | The target the system is graded on |
| Scope | Team, system, document type and date of the source document | Platform engineering wiki page, March 2026 | Collisions sit inside teams; scope narrows them |
| Deciding evidence | The words or fields that settled the sense | Link to a service repository and a review checklist | Lets a grader confirm the label from the record |
| Status on the document date | Current, renamed, retired or replaced, with any successor | Current | Stops an old document being graded against a new sense |
| Adjudication | Who labelled, who checked, and whether they agreed | Two annotators, agreed | Separates hard items from labelling noise |
Disambiguate or ask: the evidence-grounding test
Grade a system on three behaviors, not one: choosing the right sense when context decides, asking when it cannot, and not asking when the answer is on the page. A system that always asks is useless, and one that never asks is confidently wrong on exactly the items where the deciding evidence is missing.
Build the test set in three strata. Context-decided items keep the deciding evidence inside the document or its metadata. Ask items remove that evidence, so two senses stay equally plausible and the correct output is a clarifying question that names both. Dated items pair an older document with a renamed or retired term, where the correct sense is the one in force on the document date. Report each stratum separately, and treat a confident wrong answer on an ask item as the most serious error.
An Illustrative pair shows the strata at work. A platform team's ticket says the ADR for the new queue is due Friday and links a service repository; the link and the team scope decide architectural design record, so asking would be an unneeded question. A ticket from a shared intake queue says only that the ADR is due Friday, with no link and no team, and the right output names both senses and asks. Moved into a dated stratum, a ticket from before a programme was renamed tests whether the system reads the sense in force when it was written.
Asking before acting on an ambiguous tool parameter follows the same logic, and a companion page covers that case. Acronyms differ because the candidates come from a glossary the organization maintains, so the glossary's version is part of the gold label. Report these scores:
- Sense accuracy on context-decided items, with the number of candidate senses beside each result.
- Ask rate on ask items, and the count of confident wrong answers there.
- Unneeded questions on context-decided items.
- Accuracy on dated items, split by renamed and retired terms.
- Agreement between annotators in each stratum, so labelling noise is not read as model error.
Is the problem the acronym or the missing context?
Often the context, and the dataset should be able to tell which. Relevant text can be outdated, restricted or simply too thin: a sense can be right for an old document and wrong today, the page that would settle it can be one the reader may not open, and a one-line ticket may never contain the deciding words.
Keep those cases apart in the labels. Mark dated items with the status field, mark items whose deciding evidence sits in a document the reader cannot access as ask items for that reader, and mark items with no deciding evidence anywhere as unresolvable rather than forcing a guess into the gold set. Public research datasets for acronym disambiguation exist, such as the one introduced at COLING 2020, but they are built from scientific papers and carry no team scope, glossary version or access restriction, which are exactly the fields company records add.
What these counts cannot tell you
A collision in a list is not a confusion in practice. The counts measure how many short forms an organization has given more than one meaning, not how often each meaning appears in documents, so a rare sense of a common acronym weighs the same as a frequent one.
Both lists are curated by people. The NASA file merges eight source lists of very different sizes, and its readme says it expands over 25,000 NASA acronyms while the file at this commit holds 22,278 distinct abbreviation strings, so the denominator depends on how entries are counted. The Ministry list is contributed by staff and carries no revision date per entry. Normalization merges some variants and misses others, which is why both hand checks are reported. None of this measures SourceX data or any buyer's corpus.
Methods note: the files, the commits and every rule applied
Counted on 2026-10-28 by a script that reads lists/acronyms.json from nasa/NASA-Acronyms at commit 8a1235f6d96bd60e1156c75ba8076da0d8118c87 and README.md from ministryofjustice/acronyms at commit 3ad8771d710730d4dd74cc78e0eb5399e86cf57d, running no code from either repository. An abbreviation key is the abbreviation text with surrounding spaces removed, case-sensitive. A meaning key is the expansion lowercased, with parenthesized text removed and runs of other characters collapsed to one space; 15 NASA rows and 4 Ministry rows with empty meanings were dropped. A collision is an abbreviation with two or more distinct meaning keys, and a NASA collision is same-source when a single source value holds two of them. For the Ministry, a sense has a context note when the Info cell of its first row in file order is not empty.
The NASA sample was drawn by sorting the colliding abbreviations with Python's default string order and taking random.Random(789).sample of 30 in Python 3; all 76 Ministry collisions were read. The change-over-time search matched previous, previously, formerly, formally, former, no longer, replaced, replacement, renamed, old name or superseded as whole words, ignoring case. Percentages are exact fractions rounded once, half up. The hand labels and counts are kept with the script, and saved files hold acronyms and counts only.
Requesting documents where acronyms appear in context
A useful enterprise acronym ambiguity dataset pairs real documents with the organization's own glossary, dated, plus enough metadata to record scope: team, system, document type and date. Public glossaries supply candidate senses but no occurrences, which is why counts like these are a starting point rather than a test set.
SourceX licenses business data directly from companies and finds datasets for a developer's request rather than holding them in stock. The request covers scope, volume, language, format, timing and licensing, and it can ask for glossaries with their revision history and for documents that keep team and system metadata. De-identification runs through approved processing routes, and because internal acronyms can point to a named unit or programme, a request should say which acronyms must survive and which may be generalized. Put the acceptance test in the request too: for example, a minimum share of items with the deciding-evidence field filled and an ask stratum of stated size.
Frequently asked questions
What is an enterprise acronym ambiguity dataset?
It is a set of real company documents in which every acronym occurrence is labelled with the sense that applies, the glossary and version that supplied the candidates, the team and date scope, and the evidence that decided it. Items where context cannot decide are labelled as ask items, so a system can be graded on when it should request clarification.
How common are ambiguous acronyms in company documents?
Public lists suggest collisions are routine: 30.3 percent of the 22,278 abbreviations in NASA's list and 6.6 percent of the 1,151 in the Ministry of Justice glossary carry more than one meaning. Those are counts of listed meanings, not of how often each appears in documents, so measure your own corpus before relying on either figure.
Should an AI assistant ask what an acronym means?
Only when the context cannot decide. Grade it on three things separately: picking the right sense when the evidence is present, asking when it is missing, and not asking when the answer is on the page. A confident wrong answer where it should have asked is the costliest error.
Why does a sense label need a date?
Because meanings expire. In the Ministry of Justice glossary, 82 entries record a rename, replacement or retirement, so a document written before the change uses the old sense. Without the document date and the term's status, a correct label on an old document can be graded as wrong.
Can a public glossary replace labelled company documents?
No. A glossary lists candidate senses but not where each one occurs or what decided it. It is useful as the sense inventory, versioned, while the test items have to come from documents that keep their team, system and date metadata.
Sources
- NASA's public acronym list file holds abbreviation, expansion and source fields for each entry, merged from eight source lists. Source
- The NASA-Acronyms browser extension script loads lists/acronyms.json from the repository. Source
- The NASA-Acronyms readme says the tool expands over 25,000 NASA acronyms. Source
- The Ministry of Justice acronyms readme holds a table of abbreviations, definitions and context notes and says the site is intended for internal MoJ use only. Source
- The Ministry of Justice publishing script builds the acronym page by parsing the readme table rows after its header and separator. Source
- The Joint Commission created its Do Not Use list of abbreviations in 2004 to meet a National Patient Safety Goal requiring accredited organizations to keep such a list. Source
- A COLING 2020 paper introduced acronym identification and disambiguation datasets built for the scientific domain. Source
- SourceX's request page says it licenses business data directly from companies, asks for scope, volume, language, format, timing and licensing, and sources datasets on request. Source
- SourceX's how-it-works page says data is extracted, transferred and de-identified through approved processing routes. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.