Vocabulary | What Is a Corpus Provenance Record? How to Trace the Source of Every Word Count

CHAPTER 1 OF 23

1. Start with a result that must be traceable

A corpus count such as ‘the target appears nine times’ is a derived observation. To evaluate it, a reader needs to know which texts were eligible, which words were searched and whether copied or revised documents contributed.

A practical example

A school archive reports nine uses of submit. If the texts included duplicate PDF and HTML copies, nine may describe file appearances rather than independent original wording.

The tempting but incorrect conclusion

The mistake presents the number as a property of a whole community without recording source files or how the counting was performed.

A more defensible decision

Keep a result ID linked to the source set, query rule, processing version and date. A number without this lineage is difficult to verify or compare later.

Check yourself: What minimum question should a reader ask of a frequency result?

Explained answer: Which text version and query produced it, and what items were included or excluded?

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Next

CHAPTER 2 OF 23

2. Distinguish metadata from provenance

Metadata describes properties such as title, author, language, date and genre. Provenance specifically records origins, transformations and responsible processes linking source material to derived outputs. The two overlap but are not identical.

A practical example

A source page may have a publication date and author as metadata. A provenance chain adds that it was downloaded at a given time, converted to text and filtered using a particular pipeline before counting.

The tempting but incorrect conclusion

The shortcut says ‘we recorded the title, therefore the count is reproducible’ while leaving extraction and duplicate decisions undocumented.

A more defensible decision

Keep descriptive metadata and processing history together but distinguish their roles. The title describes the item; the lineage explains how the analysed representation was produced.

Check yourself: Can a title and date alone reproduce an extracted word count?

Explained answer: Usually not. Text version, extraction and search rules also matter.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 3 OF 23

3. Use source IDs that do not silently change

A stable identifier lets readers trace a derived record to the correct source without relying solely on a filename or displayed title, which may be reused. A version or snapshot identifier helps distinguish changed content over time.

A practical example

File N01 could mean the original notice PDF, while N01v2 denotes a corrected notice with a new deadline. The stable relationship should survive a renamed download or storage location.

The tempting but incorrect conclusion

The mistake uses the same unqualified ID for two different text versions, causing an old count and a corrected count to appear as though they came from the same bytes.

A more defensible decision

Assign stable source and version IDs, record relationships and preserve a lawful original when possible. If the source must be private, the internal ID can remain meaningful without exposing its content publicly.

Check yourself: Does a constant filename guarantee identical source content?

Explained answer: No. Files can be overwritten or updated under the same name.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 4 OF 23

4. Understand the W3C Entity–Activity–Agent model

W3C PROV-O describes an entity as a thing with fixed aspects, an activity as a process acting upon things, and an agent as a party responsible for an activity or entity. This vocabulary provides one standard way to express a processing lineage.

A practical example

The original notice is an entity, text extraction an activity, the extracted file another entity, and the operator or software an agent. A later query activity uses the extracted text and generates a count report.

The tempting but incorrect conclusion

The error treats a software name or date as the entire provenance record. The relationship among source, processing and output must also be represented.

A more defensible decision

Use the full triple when practical: what thing, what happened to it and who or what performed the activity. A simple table can convey the relation without requiring RDF expertise.

Check yourself: In the example, which part is an activity?

Explained answer: Extracting the words from the PDF, rather than the PDF itself or the resulting text file.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 5 OF 23

5. Track direct and indirect derivations

The W3C vocabulary distinguishes basic derivation from more specific relations such as revision and quotation. These distinctions help clarify whether text was copied, altered or independently created.

A practical example

A published PDF can derive from a draft. A later revised notice may be a revision of that PDF. A webpage reproducing a paragraph may be a quotation of the original notice.

The tempting but incorrect conclusion

The shortcut labels all related files independent new sources, inflating counts of distinct authored wording.

A more defensible decision

Record relationships such as derived from, revised from or quoted from where they are supported. Never infer a common origin solely because two short common phrases match.

Check yourself: Does substantial reuse make a new file a new independent authored wording choice?

Explained answer: Not necessarily. Its lineage and the analytic question determine how reuse is counted.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 6 OF 23

6. Separate source origin from collection event

A document’s writing date is not necessarily the date it was downloaded, indexed or analysed. These dates matter for time-based subsets and reproducibility.

A practical example

A circular issued in March may be downloaded in May and indexed in June. A report calling the source ‘June language’ because of the indexing date could misrepresent its original communication period.

The tempting but incorrect conclusion

The mistaken analysis treats an ingestion timestamp as the date the author used the word.

A more defensible decision

Store distinct fields for writing/publication, acquisition, processing and report generation, with uncertainty when original dates cannot be verified.

Check yourself: Can a download date substitute for the publication date?

Explained answer: Not automatically. They refer to different events.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 7 OF 23

7. Preserve raw and processed text separately

Extraction, punctuation normalisation and text cleaning can change token boundaries or remove content. A researcher should be able to distinguish the original evidence from the representation used for counting.

A practical example

A PDF includes a repeated footer on every page. The processed text may exclude footers, reducing eligible tokens and target hits. Both original and processed versions have valid but different purposes.

The tempting but incorrect conclusion

The shortcut edits the raw file to match the cleaned one and later claims the archive never contained the footer.

A more defensible decision

Preserve a lawful immutable raw copy or trustworthy source reference, and version derived text with documented transformations. Avoid exposing private source material without permission.

Check yourself: Does a cleaned text replace the historical original?

Explained answer: No. It is a derived representation that should retain its relationship to the source.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 8 OF 23

8. Define what content hashes can and cannot prove

A cryptographic hash can support byte-level integrity checking: identical bytes give the same digest under the same algorithm, while a changed digest signals changed bytes or processing. A hash does not establish authorship, copyright permission or that two passages mean the same thing.

A practical example

A downloaded PDF and its extracted plain text are different byte sequences and therefore have different hashes, even if the rendered text conveys the same message.

The tempting but incorrect conclusion

The mistaken conclusion says a matching file hash proves a document is true, lawfully acquired or independently authored. It only provides evidence about equality of the checked byte content under the hash algorithm.

A more defensible decision

Record the algorithm, input representation and digest when integrity checks are useful. Keep authorship and source trust as separate questions.

Check yourself: Does a cryptographic hash prove who wrote a document?

Explained answer: No. It is primarily evidence about content identity or integrity, not authorship verification by itself.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 9 OF 23

9. Show how a query becomes a derived result

A reported frequency is derived from an exact target definition, search pattern, processing conventions and eligible source set. The same source files can produce different correct counts under different rules.

A practical example

The query for submit alone excludes submitting and submitted. A lemma query may include those forms. Either result can be correct if the target and processing are stated.

The tempting but incorrect conclusion

The shortcut pastes the count into a report without saving the query and later treats a new lemma count as a direct replication of the old exact-form result.

A more defensible decision

Record query string, exact-form or lemma definition, tokenisation, filters, corpus snapshot and reporting denominator. A reviewer should be able to rerun the calculation.

Check yourself: Can the same original notice generate two correct word counts?

Explained answer: Yes, when the queries or eligible units differ and their scope is explicit.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 10 OF 23

10. Use source lineage to detect duplicates

Different files can share an underlying text. Provenance captures whether one was copied, syndicated, reformatted or revised, helping an analyst choose a consistent duplicate policy.

A practical example

One notice is in HTML and PDF, and a corrected version appears later. They are three file records but may represent one original statement and one substantive revision.

The tempting but incorrect conclusion

The shortcut collapses all three automatically or counts them all as independent without explaining the source relationship.

A more defensible decision

Build a simple origin and version register. Choose unique-text, version-aware or exposure-event counting according to the question.

Check yourself: Does a PDF and HTML copy automatically mean two independent authored texts?

Explained answer: No. They can be alternate representations of one original document.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 11 OF 23

11. Know why file location is not source identity

Websites move, URLs redirect and archives reorganise. A link can change even when a document is the same, while a stable URL can serve revised content. Source identity and access location should be distinguished.

A practical example

A notice is moved from a school webpage to an archive page without changing its content. A new URL alone does not imply a new authored document.

The tempting but incorrect conclusion

The misleading claim counts every URL as a unique writer or independent document.

A more defensible decision

Preserve original and current access locations and a stable identity relationship where evidence supports it. A URL is a locator, not a universal proof of identity.

Check yourself: Can different URLs point to copies of the same source?

Explained answer: Yes. Location and intellectual origin are different properties.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 12 OF 23

12. Construct an end-to-end lineage table

The complete toy register follows one fictional notice from raw PDF to extracted text, a cleaned version, a query report and a corrected subsequent notice. It shows both the transformation pathway and the source-revision relationship.

A practical example

Record R01 raw PDF has 500 eligible tokens and four exact submit matches. After a chosen extraction and footer exclusion, the processed version P01 has 450 eligible tokens and three matches. A report then records 3/450×1,000,000≈6,666.67 per million.

The tempting but incorrect conclusion

The mistaken report calls the rate 8,000 per million by dividing the raw four hits by the raw five hundred tokens, while labelling it as a cleaned-text result. That mixes versions or fails to identify the calculation being reported.

A more defensible decision

Tie every count and denominator to the same processed source version. Check the report result against the stated source and query. Label the examples as fictional.

Check yourself: What is the cleaned-text rate for three matches among 450 eligible tokens?

Explained answer: 3/450×1,000,000≈6,666.67 per million.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 13 OF 23

13. Preserve a revision as new evidence

A corrected source may change a date, obligation, number or lexical item. The revised version should be linked to the original rather than silently overwriting it or treated as an unrelated first draft.

A practical example

A notice changes ‘submit the form by Friday’ to ‘submit the form by Monday’. The source lineage includes a substantive deadline revision, even though the target word submit remains present.

The tempting but incorrect conclusion

The shortcut discards the original and then claims the report always said Monday. That erases the historical instruction and may distort time-based usage analysis.

A more defensible decision

Keep a revision relationship, effective date and relevant changes. Define whether your analysis counts initial publications, all versions or final approved texts.

Check yourself: Can an updated notice be a distinct communication event without being a wholly independent original?

Explained answer: Yes. It may be a revision of one source and a separate issued event.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 14 OF 23

14. Distinguish provenance from copyright permission

A provenance record can identify where material came from without granting the right to redistribute it. Source tracing, privacy safeguards, licensing and ethical access are separate requirements.

A practical example

An educational corpus may include student essays held under school policies. A public report can name the method and anonymised counts without publishing identifying raw essays.

The tempting but incorrect conclusion

The mistake assumes that recording a source URL or researcher name makes any copied text safe and lawful to republish.

A more defensible decision

Track access rights and licensing where appropriate and minimise sensitive content. Publish safe derived statistics or invented examples when source redistribution is not permitted.

Check yourself: Does transparent provenance automatically give permission to reuse personal writing?

Explained answer: No. Permission, confidentiality and data rights must be evaluated separately.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 15 OF 23

15. Connect provenance with reproducibility

A reproducible analysis lets someone with permitted access reconstruct the result from the specified source and procedure. Provenance supports that reconstruction, but alone cannot guarantee the code is correct or the data representative.

A practical example

An auditor receives the processed source ID, rule version and exact query; they obtain three matches and recompute the reported rate. A subsequent version with four matches can also be traced.

The tempting but incorrect conclusion

The shortcut treats a complete-looking metadata form as proof of correctness without checking the code, duplicate policy or statistical interpretation.

A more defensible decision

Pair provenance records with sample manual validation, script versions and source coverage notes. Distinguish provenance, reproducibility and methodological validity.

Check yourself: Can a perfectly documented process still be methodologically wrong?

Explained answer: Yes. It may consistently process an unrepresentative or misclassified dataset.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 16 OF 23

16. Account for source clustering and independence

A corpus provenance graph can reveal that many documents share an author or template. This information helps prevent a frequency chart from being interpreted as independent evidence from many writers.

A practical example

Ten related notices may all derive from one institutional template. Their appearances can reflect repeated exposure while independent authoring remains concentrated.

The tempting but incorrect conclusion

The mistaken report translates ten files into ten separate people using the expression.

A more defensible decision

Preserve writer, institution and template relationships where lawful and relevant. State whether the statistic concerns tokens, unique texts or independent sources.

Check yourself: Do ten related file records necessarily represent ten independent writers?

Explained answer: No. Shared provenance may show common authorship or copied material.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 17 OF 23

17. Use versions to explain apparent trends

If a later corpus count differs, it might reflect new documents, corrected metadata, changed extraction or a revised query. A provenance record helps isolate those causes before attributing the difference to language change.

A practical example

Suppose a rate drops after the analyst removes repeated footers. No author has necessarily changed writing habits; the effective token set changed.

The tempting but incorrect conclusion

The shortcut calls the difference a decline in students’ vocabulary use when it is a processing revision.

A more defensible decision

Compare source and pipeline versions, then check which observed language events actually differed. Label methodological changes separately from historical usage claims.

Check yourself: Can frequency change without writers changing their language?

Explained answer: Yes. Processing, inclusion and duplication rules can change the count.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 18 OF 23

18. Define provenance resolution appropriate to the task

Not every classroom project needs a complex graph. A small corpus can often use a table containing source ID, version, collection date, transformation, query and result. More consequential projects may require fuller event histories and validation.

A practical example

A tutor’s fictional ten-sentence corpus can record one stable source label and a handwritten query rule. A research archive of thousands of copyrighted documents may need controlled identifiers and structured metadata.

The tempting but incorrect conclusion

The mistake assumes either that provenance is only for advanced software engineers or that one title column is always sufficient.

A more defensible decision

Match granularity to the claim’s importance and future use, but retain enough information to reconstruct the result. Plain-language tables can be rigorous.

Check yourself: Does a useful provenance record require specialised RDF software?

Explained answer: No. A clear source-and-transformation register can communicate the essentials.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 19 OF 23

19. Avoid attaching fabricated metadata to real sources

When publication dates, authors, licences or source lineages are unknown, a provenance record should say unknown or unverified rather than filling fields from guesses. A false attribution can undermine the whole audit.

A practical example

A copied snippet arrives without an original URL. The analyst might recognise the topic but still cannot truthfully identify the original author without evidence.

The tempting but incorrect conclusion

The shortcut invents an author and source date to make the spreadsheet look complete.

A more defensible decision

Use explicit unknown values, seek authoritative documentation when necessary and distinguish observed file properties from inferred origin.

Check yourself: Is an incomplete honest record better than a complete invented one?

Explained answer: Yes. Unknown provenance should be marked rather than disguised as established fact.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 20 OF 23

20. Give readers a traceability chain

A useful published table allows a reader to move backward from report number to query, processed file, extraction and original source; and forward from original text to any derived analyses. The direction of tracing matters in corrections.

A practical example

Start from Q01: three hits in P01. Follow P01 to E01, the extraction and cleaned representation of raw R01. A revised notice R02 forms a different branch in the lineage.

The tempting but incorrect conclusion

The mistake presents a diagram with arrows but does not specify whether an arrow means copy, revision, query use or an authored derivative.

A more defensible decision

Label relationships with verbs such as derived from, revised from, processed by and generated by. This keeps the graph meaningful.

Check yourself: What should be traceable from a frequency report?

Explained answer: Its source content version, processing rules, exact query and denominator.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 21 OF 23

21. Apply provenance to reading and school work

Students also need provenance literacy. A quotation in History, an image caption, a scientific figure and a statistical claim all require evidence of origin and interpretation. Vocabulary terms such as ‘original’, ‘revision’, ‘copy’ and ‘derived’ prevent careless conclusions.

A practical example

A History essay compares two reproduced newspaper accounts. Knowing one is a reprint of the other changes how many independent accounts support the claim.

The tempting but incorrect conclusion

The shortcut treats two web pages as corroboration without checking whether both reproduce the same original report.

A more defensible decision

Teach students to check who created the information, when it was made, what was changed and whether apparent corroboration is independent.

Check yourself: Why might two webpages not provide two independent historical sources?

Explained answer: They may both derive from one original article or document.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 22 OF 23

22. Audit a source record before publication

Before citing a corpus statistic, confirm source identity, rights, version, extraction, duplicate policy, query, token denominator, reported result and revision history. Keep the record small enough to inspect but complete enough to recreate.

A practical example

An analyst who can reproduce 3/450×1,000,000 from the documented cleaned text has a transparent toy result. It remains a local count in a fictional notice, not evidence about national writing.

The tempting but incorrect conclusion

The mistake gives a precise-looking number with no traceable inputs, or uses a real source name to legitimise invented data.

A more defensible decision

Publish the audit and its limitations. Where raw texts cannot be shared, provide lawful aggregated data and enough procedural detail for authorised review.

Check yourself: What is the minimum quality check for a derived frequency count?

Explained answer: Confirm that numerator and denominator come from the same declared data version and query.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Next

CHAPTER 23 OF 23

23. Finish with the claim’s correct scope

Provenance makes language evidence understandable but cannot replace valid sampling, semantic interpretation or trustworthy sourcing. A documented bad estimate is still a bad estimate, though its weaknesses become easier to locate.

A practical example

A modest conclusion reads: ‘The fictional processed notice P01 contains three exact target hits among 450 eligible tokens after specified cleaning.’ It does not claim to describe all notices or every pupil.

The tempting but incorrect conclusion

The overclaim says an invented example proves institutional or national vocabulary patterns. No population sample or empirical observation was collected.

A more defensible decision

Name what the source actually supports, identify the processing path, and stop where the evidence stops. This is a stronger form of clear communication than asserting certainty.

Check yourself: What is the practical goal of provenance?

Explained answer: Make the origin, transformations and limits of a result traceable enough to support verification and correction.

A sound vocabulary method should allow another reader to reconstruct the conclusion from the same source and instruction. Record the unit of analysis, the rule and its version, and the evidence used to decide a difficult case. When a change improves a rule, identify which earlier records require review instead of silently pretending that all annotations were produced under the same conditions. When evidence remains insufficient, a qualified conclusion is preferable to an invented certainty.

All chapters · Previous · Worked register

Worked register: follow the evidence row by row

This constructed miniature register describes the lifecycle of a fictional notice. A raw PDF R01 contains a repeated footer; P01 is its processed running-text representation. A revised notice R02 is an authored update rather than an exact copy. Numbers are invented solely to make the chain and counting denominator visible.

Record ID	Entity or activity	Derived from / uses	Recorded content/count	What readers must know
R01	Raw PDF source	Original authored notice	500 word tokens; 4 exact submit hits	Source edition 1; includes footer
E01	Text-extraction activity	Uses R01	Generates extracted text T01	Transformation, extraction settings and operator recorded
P01	Cleaned processed text	Derived from T01	450 eligible tokens; 3 exact submit hits	Footer omitted under declared rule
Q01	Exact-word query and report	Uses P01; generated by query step	3/450×1,000,000 = 6,666.67 per million	Not the same denominator as R01
R02	Revised raw notice	Revision of R01	Deadline Friday changed to Monday	Track revision; no count claimed for R02
L01	Source lineage register	Tracks R01,E01,T01,P01,Q01,R02	One original, one processed line, one revision branch	Distinct entities and events have different roles

Raw R01 contains four exact target hits within five hundred word tokens, for 4/500×1,000,000=8,000 per million if the raw source is analysed as such. That is a valid arithmetic result for the fictional raw version, not for the cleaned representation.

Processed P01 contains three hits in 450 eligible tokens after the declared footer exclusion. Its rate is 3/450×1,000,000≈6,666.67 per million. The changed rate reflects a changed text representation, not a real change in the author’s communication.

Q01 explicitly uses P01 and therefore must preserve that source ID and tokenisation policy with its numerator and denominator. Combining R01’s four hits with P01’s 450 tokens would yield 8,888.89 per million, an incoherent hybrid result.

R02 is a revision because an instruction changed from Friday to Monday. The record does not supply a new token count or a new frequency result for R02. A responsible analysis would process R02 separately before claiming its new rate.

The lineage distinguishes the original source entity from transformations and reports. A processing timestamp, source publication date and later revision date should remain separate. This example does not establish any facts about real school notices.

Twelve transfer cases and reasoned responses

These fictional cases are arranged to reveal ordinary weaknesses in methods: unknown sources, renamed labels, ambiguous examples, copied content, changed versions and interpretation that outruns the data.

1. PDF and HTML

Situation. One notice exists in two formats.

Why it matters. File identities differ from authored source identity.

Reasoned response. Link alternate representations to common origin.

Further question. Is the task unique authorship or reader exposure?

2. A new deadline

Situation. A notice changes Friday to Monday.

Why it matters. Near-identical text may carry a significant revision.

Reasoned response. Keep an edition relationship.

Further question. When did the revision take effect?

3. Unknown writer

Situation. An extract lacks clear authorship.

Why it matters. A filename is not reliable attribution.

Reasoned response. Mark author unknown and seek evidence.

Further question. Which source establishes identity?

4. Footer cleaning

Situation. A footer is removed before counting.

Why it matters. The eligible corpus and rates change.

Reasoned response. Save original and cleaned versions and the rule.

Further question. Did numerator and denominator both update?

5. Two sources echo one

Situation. A page republishes a press release.

Why it matters. Apparent corroboration is not independent.

Reasoned response. Record quotation or derivation links.

Further question. How many independent authors chose the wording?

6. Stale download

Situation. A page updated after it was indexed.

Why it matters. Acquisition and publication dates differ.

Reasoned response. Capture source snapshot and event dates.

Further question. Which edition did the query use?

7. Same filename

Situation. A PDF is overwritten with new bytes.

Why it matters. Old results no longer point to one fixed edition.

Reasoned response. Use version IDs and integrity records where feasible.

Further question. What content was analysed?

8. Sensitive essays

Situation. Student writings are in a protected corpus.

Why it matters. Source traceability does not grant publication rights.

Reasoned response. Use access control and safe aggregate reporting.

Further question. Can identifiers reveal a child?

9. Lexical expansion

Situation. Exact submit search becomes a lemma search.

Why it matters. The target changed, not necessarily the source.

Reasoned response. Version and label the query.

Further question. Which word forms now qualify?

10. Format conversion

Situation. HTML is converted to plain text.

Why it matters. Navigation or punctuation may alter tokens.

Reasoned response. Record extraction and segmentation.

Further question. Does the token denominator differ?

11. Unknown licence

Situation. A downloaded article has unclear reuse rights.

Why it matters. Source location does not confer redistribution permission.

Reasoned response. Mark licence status and avoid unauthorised publication.

Further question. What licence or consent applies?

12. National overclaim

Situation. A rate from one notice is described as national.

Why it matters. Provenance cannot repair an invalid sampling frame.

Reasoned response. State the source-specific count and collect relevant evidence.

Further question. What domain was actually observed?

Eight practice questions with explanations

Try each question before reading the answer. For every revised label, say what changed; for every source record, say which entity, transformation and version led to the result. The aim is independent reasoning rather than copying a correction.

Exercise 1: Trace a cleaned rate

Question. Which record supplies the 3/450 count?

Explained solution. Processed text P01 and its exact query Q01; the rate is approximately 6,666.67 per million.

Common wrong turn. Using raw R01’s hits or token count would describe a different version.

Exercise 2: Compare raw and processed

Question. Compute R01’s raw rate with four hits among 500 tokens.

Explained solution. 4/500×1,000,000=8,000 per million, specifically for the raw source.

Common wrong turn. Do not call it the cleaned P01 rate.

Exercise 3: Identify an activity

Question. Which part is the extraction activity rather than an entity?

Explained solution. E01, the step using R01 to produce extracted text T01.

Common wrong turn. R01 and T01 are source and derived entities, not the extraction operation.

Exercise 4: Recognise a revision

Question. A deadline changes between two notices. How should they be linked?

Explained solution. R02 is a documented revision of R01, with its changed text and issue date recorded.

Common wrong turn. Treating the new notice as an unrelated original erases lineage.

Exercise 5: Explain duplication

Question. An identical HTML copy and PDF have two URLs. Do they prove two independent authors?

Explained solution. No. They can be alternate representations of one authored notice.

Common wrong turn. URLs locate files; they do not guarantee unique authorship.

Exercise 6: Separate dates

Question. Can the date a corpus was downloaded establish when a letter was written?

Explained solution. No. Acquisition and original publication or writing are different events.

Common wrong turn. Conflating them can distort chronological language claims.

Exercise 7: Inspect missing metadata

Question. The source author is unknown. What should the register say?

Explained solution. Unknown or unverified, with a plan to seek a reliable source if needed.

Common wrong turn. Guessing a plausible person undermines the provenance record.

Exercise 8: Report scope

Question. Rewrite ‘This proves schools use submit often’ using the toy result.

Explained solution. ‘In this fictional processed notice, the exact form submit occurs three times among 450 eligible tokens under the stated rule.’

Common wrong turn. A constructed single-document example does not support a population claim.

Frequently asked questions about corpus provenance record

What is corpus provenance?

The documented origin and transformation history of source texts and their derived language results.

How does provenance differ from metadata?

Metadata describes items; provenance additionally records where they came from and the processes that generated later versions or results.

Why retain source versions?

Text edits and processing changes can alter counts without a new linguistic phenomenon.

What is an entity in W3C PROV?

An identifiable thing such as an original source document, processed dataset or reported chart.

What is a PROV activity?

A process using or generating entities, such as extraction, cleaning, annotation or frequency analysis.

What is a PROV agent?

A person, organisation or software agent bearing responsibility for an activity or item.

Is a revision the same as an exact copy?

No. A revision includes meaningful changes while retaining a lineage relationship.

Can a content hash establish authorship?

No. It helps check byte-level identity or integrity, not who wrote the content.

Why record the query string?

It determines what matched and makes frequency results reproducible.

Can a result be reproducible but wrong?

Yes. A consistent computation may still use poor source coverage or inappropriate interpretations.

Does a link prove a source is original?

No. A URL is an access location, not proof of independent authorship.

Can a provenance record be a simple table?

Yes, if it conveys source identity, version, transformations, query and result sufficiently.

Must private sources be published?

No. Protect sensitive data while sharing permitted aggregate methods and counts.

Why distinguish source time from download time?

The source’s communication date and the corpus’s acquisition event answer different questions.

What should a vocabulary report cite?

The specific source and processing version, exact query, eligible denominator and any material limitations.

Compact reference glossary

• Provenance: The origin and derivation history of an item or result.
• Entity: An identifiable source, data file or derived output.
• Activity: An event or process transforming or using an entity.
• Agent: A person or system responsible for a process.
• Derivation: A relationship between an output and source entity.
• Revision: A new version of an existing source with changed content.
• Quotation: Repetition of some or all of an earlier source.
• Metadata: Descriptive properties about a resource.
• Extraction: Converting source content into an analysable representation.
• Tokenisation: Dividing analysed text into countable elements.
• Content hash: Digest supporting checking of the bytes of a file.
• Lineage: A traceable path of source, processing and result.

How to teach this idea without technical overload

Begin with a real communication task before introducing the technical term. For example, ask pupils why the same written word can receive different grammatical labels in different sentences or why a frequency chart changes after a copied source is removed. Let them state a prediction, then reveal which record or rule makes the difference.

Next, make a small transparent register. Students should be able to read the original sentence or source title, follow its classification or transformation, and reconstruct one summary number without a hidden spreadsheet. Ask a second reader to check whether the stated rule has been applied consistently.

Third, deliberately introduce one difficult example: a negation, a changed file version, an ambiguous sentence, or a mislabeled source. Require learners to revise the conclusion in words, not simply change the answer letter. This helps them see that vocabulary precision depends on the connection between language and evidence.

Finally, test transfer to a new context rather than repeating the same exact example. A student who can identify an undocumented revision in a History source or explain a changed definition in a Science report has moved beyond memorising terminology. Record the correct reasoning and any remaining uncertainty.

Research and methodological sources

• W3C PROV-O — Entity, Activity, Agent, derivation and revision vocabulary
• CLARIN Metadata Guidance — source description, content and sensitive data
• Universal Dependencies — dataset and annotation versioning practices
• eduKateSG — Corpus Metadata

The cited sources establish frameworks and examples of the relevant professional methods; the miniature data and arithmetic above are original illustrations.

Continue through related vocabulary guides

• Vocabulary | What Is an Annotation Guideline?
• Vocabulary | What Is Annotation Adjudication?
• Vocabulary | What Is an Annotation Confusion Matrix?
• Vocabulary | What Is Corpus Metadata?
• Vocabulary | What Is Corpus Source Clustering?
• Vocabulary | What Is a Corpus Sensitivity Analysis?
• Vocabulary Learning Hub
• Vocabulary Article Library

Return to contents · Vocabulary Learning Hub