Vocabulary | What Is Corpus Metadata? How Source Details Change What Word Evidence Can Tell Us

Full chapter contents

Imagine finding the sentence “You must submit the form today.” You can inspect its vocabulary and grammar without knowing anything else. You can identify must, ask what it expresses and notice the deadline. Yet you cannot safely decide whether it is a natural model for a friendly invitation, an examination instruction or a fictional argument. That decision needs information about the situation in which the sentence occurred.

Corpus metadata is structured information describing a language collection or its components. It may tell us where a text came from, what kind of text it is, when it was produced, how it was collected and what processing it has undergone. A corpus is a collection of language material assembled for investigation. Metadata helps us understand which investigation the material can support.

In a small reading project, metadata might be a table beside six text files. In a research corpus, it may appear as searchable fields, documentation and structured headers. The practical purpose is the same: keep the words connected to enough information about their source and preparation for another person to interpret them responsibly.

The British National Corpus description offers a real example: its texts include contextual, classification and bibliographic information in headers. The project distinguishes spoken and written material and documents the collection’s scope. Those descriptions matter when choosing evidence for a language question.

Change the question, and the useful metadata changes

If you want a model for a school announcement, text purpose and audience may matter most. If you want to investigate a change over decades, the date and its meaning become essential. If you want to compare spontaneous conversation with edited prose, the source medium and editing history matter. One attractive example cannot answer all three questions equally well.

A helpful first habit is to finish this sentence: “I can use this example to investigate ___ because I know ___ about its source.” If the second blank is empty, you have found the next research task. You may still read the example, but your conclusion should remain narrower.

Contents · Next chapter

Three layers often appear together in a corpus interface. Keeping them distinct prevents an important kind of confusion: treating a description of the source as if it were a property of every word, or treating an analyst’s interpretation as if it were part of the original text.

The sentence is the material being studied. The metadata records its provenance and context. The annotations add a linguistic analysis. A field called request might be an annotation in one project and a document-level classification in another. The useful distinction therefore concerns what the field describes and how its value was assigned, rather than a universal rule about the field’s name.

The American National Corpus’s first-release encoding guide illustrates this separation. Its headers describe provenance and contextual categories, while linked annotation documents identify token boundaries and linguistic information such as parts of speech and lemmas. This is one documented project design, not a compulsory file format for schoolwork.

Do not let a label decide the answer in advance

Suppose a text is labelled formal. That does not mean every sentence contains formal vocabulary. A formal report can quote an informal conversation. If you count the quoted words without noticing their role, you may describe the author’s usual style incorrectly. A document-level label is useful context; the surrounding passage still needs reading.

Likewise, a tool may label record as a noun when it is functioning as a verb in a particular sentence. Metadata cannot repair that analysis automatically. You need to inspect the linguistic annotation and the text. When a result seems surprising, ask which layer might be responsible: the words themselves, the source description or the interpretation added later.

This three-layer check is useful even without software. Put the sentence on one card, its source note on a second and your interpretation on a third. You can revise the interpretation while keeping the original wording and source information intact.

Contents · Previous chapter · Next chapter

A beginner can easily turn metadata into an endless form. There is always another detail that might be interesting. A better approach is to begin with the decision the learner wants to make and identify which differences could change that decision.

Consider the question “How can I make instructions easier to follow?” Useful material might include several authentic instructions that readers can understand and use. Useful metadata could include task, intended audience, medium, source and whether the extract is complete. The precise age of the author would usually add little to this classroom question. Collecting it would create work without improving the comparison.

Now consider “Do the same words organise a story and a set of instructions?” Text type, length and a consistent rule for counting the target words become central. You should also know whether the texts were independently written or adapted from a shared template. A large number of copied examples would not provide the same variety of evidence as independent sources.

Write a one-sentence collection contract

For a small project, write: “We will compare the use of next in two teacher-created notices and two teacher-created stories, using equally sized extracts and the same counting rule.” This statement identifies the material, the comparison and the limits. It does not pretend that four invented passages represent all English.

Under that contract, you need an identifier for each extract, its text type, its length, its source status and the counting procedure. You do not need a complicated research database. If the question later changes to real-world usage, the material and its source records must change too. Renaming the file “authentic corpus” would not transform constructed examples into observations of independent language use.

Write down what would count as a useful result. For this project, the result could be a more precise explanation of the different functions of next. It need not be a large frequency claim. This keeps the exercise connected to vocabulary learning rather than to producing an impressive-looking spreadsheet.

A good field earns its place by helping selection, interpretation, verification or responsible reuse. If you cannot explain its job, leave it out initially. You can add a justified field later while documenting the change. Thoughtful simplicity is more useful than a long form filled with guesses.

Contents · Previous chapter · Next chapter

The following record is a teaching design for a small classroom collection. It is not a universal standard. Each field answers a practical question, and the collection contract determines whether more detail is needed.

The identifier is especially useful. A title can change, and two files can share a title. A stable sample ID allows you to say that N01 was corrected without confusing it with N02. The ID does not prove that the sample is accurate or independent; it simply helps people refer to the same object consistently.

Make missing information explicit

Do not turn an empty date cell into an estimated year because the table looks untidy. Use an agreed value such as unknown, and explain what it means. Distinguish not recorded from not applicable. A recording date may be unknown for a spoken extract, while it is not applicable to a written note that was never recorded.

A note can also be provisional. “Probably a classroom instruction; classification not checked” conveys a different level of certainty from “Instruction, confirmed from the source page.” If a filter requires confirmed instructions, those records should not silently enter the same group.

Finally, try a handover test. Give the record and extract to another learner. Can that person identify the sample, understand what was counted and repeat the selection? If they need to ask what most columns mean, improve the field descriptions before collecting more material.

Contents · Previous chapter · Next chapter

These labels are useful because a word’s job depends partly on the activity around it. They can also become a source of confusion when a project uses them as interchangeable shortcuts. For a classroom investigation, define the categories in ordinary language before applying them.

You might use genre for a recognisable kind of text, such as a recipe, news report, short story or meeting agenda. You might use register for the language associated with a situation of use, considering purpose, participants and medium. Researchers use these terms in different frameworks, so report your local definitions rather than suggesting that one simple table settles the terminology for every discipline.

Topic concerns what the text is about. Audience concerns whom it addresses. A story and a set of instructions can both concern a school garden. Two instructions can address different readers: a beginner preparing a pot and an experienced groundskeeper checking irrigation. Their shared topic does not make their wording directly interchangeable.

A four-question classification check

For an invented notice beginning “Before planting, loosen the soil,” ask: What kind of document is this? What is it trying to help someone do? What is it about? What does it assume the reader already knows? The answers might be instruction sheet, plant a seed, gardening and basic familiarity with a pot and soil. Each answer contributes a different piece of context.

Now compare an invented story: “Before planting, Mira remembered her grandfather’s advice.” The phrase before planting is shared, but its role differs. In the notice it helps organise a procedure. In the story it locates an event and introduces a memory. Metadata points towards that difference; close reading explains the particular use.

A common repair is to split an overloaded column. If one cell says “formal science children,” three dimensions have been packed together. Separate topic, intended audience and the project’s style classification. This makes it possible to compare texts about the same topic while keeping their purposes visible.

Resist ranking the groups as inherently better or worse. Conversational language is not a defective version of a written report. A useful vocabulary choice fits its communicative purpose. The investigation should help learners choose deliberately across situations, rather than make them replace every familiar expression with a more elaborate one.

Contents · Previous chapter · Next chapter

A date looks precise, but its meaning may be unclear. A digital file can have several relevant dates: when the words were composed, when they were spoken, when the source was published, when it was scanned, when the transcript was prepared and when the file entered the collection. These are different events.

Suppose a story was published in 1985, scanned in 2024 and added to a class collection in 2026. A field containing only “2026” could lead someone to call its language contemporary. The number is not necessarily false; the field is insufficiently described. The repair is to name the event: publication year 1985, digitisation year 2024, collection year 2026.

Match the date to the claim

If your question concerns vocabulary in published stories during a period, publication date may be an appropriate selection criterion. If your question concerns a speaker’s language at a particular moment, the recording date is more relevant. If your question concerns changes introduced by a later editor, you may need both the original and revised editions with their respective dates.

Do not imply more precision than the source supports. “Approximately the late 1990s” should not become “1 January 1998” merely because software prefers a full date. For a simple learning project, a clear text note is often safer than a precise-looking invented value.

A reprint creates another question. Its publication year may differ from the year of the wording it reproduces. A revised edition may change some language while retaining most of the earlier text. Record what you know about the edition, and limit historical claims if that relationship remains uncertain.

The useful vocabulary here is created, published, recorded, transcribed, revised and collected. Learning these distinctions improves the metadata itself. It also helps learners read source notes in history, literature and research without collapsing several stages into one unexplained date.

Before drawing a time-based conclusion, pick one sample and trace its dates aloud. If the chain cannot be explained, investigate the record or choose a question that does not depend on the missing date. An uncertain historical comparison can still become a useful lesson in reading and describing sources.

Contents · Previous chapter · Next chapter

Corpus metadata sometimes describes authors, speakers, settings and language varieties. Such information can support carefully designed questions, but it can also invite assumptions. A text’s place of publication is not automatically its author’s home, and a speaker’s recording location does not establish their entire language history.

For a school vocabulary exercise, begin with information about the communicative event. Was this an interview, a classroom explanation, a conversation among friends or an edited article? Was it written for a local audience or a broader readership? These descriptions may explain the wording more directly than a detailed personal profile.

Separate the source of a description from the description

Imagine a record labelled “Singapore English.” Ask what that means within the collection. It could indicate the publication’s location, the recording setting, the project’s language-variety category or a participant’s own description. Those are not identical. A clear field definition protects the reader from treating a convenient label as a complete biography.

If a source does not state a person’s background, do not infer it from their name, appearance or a single pronunciation. Use unknown or omit the field when it is not relevant. Guessing can create misleading comparisons and unnecessary personal information.

Languages can also be mixed within a text. A document labelled English may contain a quotation, name or conversation in another language. The document-level label helps locate material, but it does not explain every item inside it. A vocabulary investigation may need a note identifying the specific passage being studied.

When working with children, a public or teacher-created example set is often enough to teach the method. There is no need to publish classmates’ names, family details or private messages. Use neutral sample IDs and collect only information that the learning question actually requires.

The educational aim is to understand variation without turning group labels into judgements about people. A pattern in a selected collection describes that evidence under its collection rules. It does not establish what every person in a community says, knows or should sound like. Careful metadata makes those limits visible before a claim travels beyond the classroom.

Contents · Previous chapter · Next chapter

A corpus may contain complete documents, extracts from documents or several pieces drawn from the same source. If you call every file an independent text, you may overstate the variety of the collection. Metadata should help you see how the pieces are related.

Suppose a class collects twelve instruction sheets. Six came from the same template, with only the activity name changed. The class has twelve files, but it does not have twelve independently developed ways of giving instructions. The repeated wording might be useful evidence about the template. It is weaker evidence of broad agreement across unrelated authors.

A source map can be simple

Record a sample ID for the extract and a source ID for the larger document or shared template. Two samples can have different sample IDs while sharing one source ID. If the collection includes revised editions, add an edition note. If a website republishes another source, record that relationship when known.

This does not mean every similarity is a duplicate. Two writers may independently use a conventional phrase such as “For more information.” The decision to exclude or group material should follow a declared rule and relevant evidence. Do not delete ordinary repeated language merely because repetition is the feature you are studying.

The original BNC description explains that its written material includes samples as well as complete shorter texts. Its sampling design is therefore part of how a user should understand the collection. A classroom project need not adopt that design, but it should state its own.

For a small set, ask three questions: How many files are present? How many source documents do they represent? How many independent sources or authors are relevant to the claim? The answers may differ. A spreadsheet should not hide that distinction behind a single total labelled texts.

Vocabulary helps here too. Extract, edition, copy, revision and source name different relationships. When a result seems unusually strong, check whether the same passage has entered several times under different filenames. Correcting that bookkeeping problem may change the interpretation more than adding another hundred examples.

Contents · Previous chapter · Next chapter

The following dataset is entirely invented for teaching. The counts are stipulated so that we can practise the reasoning; they are not findings from a research corpus. Imagine four teacher-created extracts, each containing 500 word tokens under one agreed counting rule. We search for the exact word form next, ignoring capitalisation.

The two instruction extracts contain 10 occurrences in 1,000 tokens. The two stories contain 2 in 1,000 tokens. Because the group lengths are equal, the raw totals already make the rate comparison straightforward: 10 versus 2 occurrences per 1,000 tokens in this constructed set.

What the metadata lets us do

The purpose field allows us to form the two groups. The token counts show that unequal length is not causing the difference here. The source note tells us that these are constructed examples, so the result cannot establish the normal frequency of next in English instructions or stories.

Now inspect the sentences. Stipulate that eight instruction occurrences organise a sequence, while two locate something beside another object. In the stories, one occurrence marks sequence and one marks location. This additional classification is an analysis of word use, not a new fact about the source type.

The narrower comparison of sequencing uses becomes eight versus one per 1,000 tokens. The first count answered “How often does this written form occur?” The second answered “How often does this form perform the selected function?” Both are legitimate under the stated rules, but they are different questions.

Turn the analysis into a writing decision

A learner revising instructions might now inspect whether each next clarifies the order of actions. A learner revising a story might ask whether repeated sequencing words make the narrative mechanical. Neither learner needs to maximise or minimise the count automatically. The purpose is to understand what the word contributes in the passage.

A responsible conclusion is: “In our four constructed extracts, next occurs more often in the instruction group. Reading the examples shows that most instruction uses organise a sequence.” An unsupported conclusion would be: “Good instructions always use next five times as often as stories.” The toy figures do not establish a standard for good writing.

For a real investigation, replace the constructed material with appropriately selected sources and document the selection. Keep the separation between group metadata, counting procedure and interpretation. The Corpus Comparison guide develops that broader comparison task.

Contents · Previous chapter · Next chapter

A student wants to investigate whether an expression appears in older and newer local notices. They find four fictional archive records. Each extract contains 1,000 tokens, and the stipulated counts concern the same exact expression. The important problem is hidden in the date columns.

If the student groups by upload year, three samples appear in 2025, with 14 occurrences across 3,000 tokens. They might describe the expression as common in recently published notices. That wording would confuse the archive’s activity with the source’s original publication.

Grouping by original publication year instead gives 14 occurrences in 2,000 tokens for the two older samples, or 7 per 1,000. The two newer samples contain 2 in 2,000, or 1 per 1,000. The arithmetic is simple; identifying which date answers the question is the central reasoning task.

What still remains unknown?

The four records do not establish a historical trend. We have not shown that the notices have comparable purposes, audiences or selection procedures. Perhaps the older texts concern a recurring activity that invites the expression, while the newer ones concern something else. The date repair improves the comparison without completing it.

The learner should also ask whether the uploaded versions reproduce the original wording. A transcription may have modernised spelling or corrected punctuation. If those features matter to the question, the processing note becomes relevant. An archive date and a source date are necessary clues, not substitutes for examining the source.

A suitable report says: “The older publication group has a higher rate in these four fictional records. Upload year would have grouped the material incorrectly for our historical question. A real claim about change would require a larger, suitably comparable collection and clear edition information.”

This case teaches a transferable reading habit: attach a date to an event. The same habit helps when comparing editions of textbooks, interpreting scanned newspapers or discussing when an online quotation was first written.

Contents · Previous chapter · Next chapter

Imagine a teacher-created audio dialogue and two written versions. The first is a transcript that retains the speakers’ words, including repetitions. The second is a polished summary written for a newsletter. Both concern the same conversation, but they do not offer the same evidence about its vocabulary.

In the invented dialogue, one speaker says: “Well, I thought we could, well, start beside the gate.” The transcript retains both occurrences of well. The summary says: “The group proposed starting beside the gate.” A learner searching the summary would find no well, but that would not show that the speaker avoided it.

Record the transformation

A useful record distinguishes original mode, current representation and processing. Original mode: spoken dialogue. Representation: transcript or summary. Processing: repetitions retained, or content condensed and paraphrased. Without these fields, a single label such as written file hides the relationship between the data and the event.

The transcript may itself be selective. It could omit pauses, laughter or overlapping speech. Those choices are not necessarily mistakes. Their suitability depends on the question. A study of word choice may not need every timing detail, while a study of turn-taking might need information that a simple transcript does not preserve.

Likewise, a polished summary can be excellent evidence for how a newsletter presents a conversation. It is simply different evidence from a verbatim account of the original speakers’ wording. The right response is to describe the transformation and choose the appropriate source.

Use the case to improve listening and writing

Ask learners to produce both a transcript and a summary from a short teacher-created recording. Then have them annotate what changed: repeated words, incomplete starts, pronouns, ordering and explicit explanations. The exercise shows that accurate summarising may preserve the main message while changing the linguistic form.

When presenting the work, label each version visibly. Do not call the summary “what the speaker said” if that suggests exact wording. Use a description such as “summary of the speaker’s proposal.” This is a vocabulary lesson about reporting relationships as well as a lesson about corpus preparation.

A final check is to ask which version would answer each question: How often did the speaker use well? How was the proposal communicated to newsletter readers? Why did the summary omit a repetition? Each question directs us to a different part of the evidence.

Contents · Previous chapter · Next chapter

Filtering feels like a neutral action: choose a value and inspect the matching texts. Yet a filter can change the collection substantially when some records are incomplete. The excluded material may matter to the question.

Imagine a fictional collection of ten instruction sheets. Six have their intended audience recorded as beginners. Four have no audience information. Searching only the beginner-labelled group returns six sheets. It does not establish that the remaining four were written for advanced readers.

Keep the excluded group visible

Suppose the target expression occurs 12 times across the six labelled sheets, totalling 3,000 tokens. It occurs 16 times in the four unlabelled sheets, totalling 2,000 tokens. The labelled group has 4 occurrences per 1,000 tokens; the unlabelled group has 8. Those figures do not let us compare beginners with advanced readers, because the second audience is unknown.

A mistaken report might say, “Advanced instructions use the expression twice as often.” The correct repair is to rename the group honestly. We have a comparison between records labelled beginner and records with missing audience information. Whether the missingness reflects a different audience, a different source or inconsistent cataloguing remains unresolved.

A practical filtering note should state the rule, the number of samples retained, the number excluded and the reason for exclusion. For a small class project, a sentence is sufficient: “We analysed six beginner-labelled sheets and set aside four whose audience was not recorded.” This allows the reader to understand the scope without reconstructing the search.

Do not repair absence with a convenient assumption

There are several reasonable next steps. Check the source to recover the missing information. Retain an explicit unknown group. Change the question so that audience is no longer required. Or limit the report to the six documented samples. Which option is best depends on the time available and the purpose of the investigation.

What you should not do is fill every blank with the label that makes the comparison easiest. A complete-looking table can be less informative than an incomplete but truthful one.

The vocabulary distinction is between absent, unknown, excluded and not applicable. These words help learners describe what the evidence does and does not contain. They are small tools for avoiding large mistakes.

A zero occurrence count means none were found under the stated counting rule; an unknown count means the count has not been established.

Contents · Previous chapter · Next chapter

A metadata table becomes more dependable when its labels come with short definitions. The definitions need not be elaborate. They should tell another reader what each field describes, which values are permitted and what to do with an uncertain case.

Consider a column called source. One learner enters the website address. Another enters the author’s name. A third enters “story.” All three are supplying something related to the source, but the column no longer supports a consistent search. Rename and separate the jobs: source location, creator if known and text type.

Give an example and a boundary

For a field called instruction, define the working criterion: “The text primarily directs the reader through an action or procedure.” Add a positive example, such as a planting guide, and a contrasting case, such as a story in which a character gives one instruction. Explain whether mixed texts receive several labels, a dominant-purpose label or a separate mixed category.

These rules do not reveal the single true classification of every text. They make the project’s choices inspectable. A reader can understand how the categories were formed and decide whether they fit a different question.

The Language Bank of Finland’s Korp format documentation provides a technical example of explicit structure: it distinguishes token attributes from structural attributes and describes how metadata is attached to text structures. A classroom table can use the same broad lesson of declaring what a field belongs to, without adopting the technical format.

Test the definitions on two easy cases and one awkward case before applying them to the entire collection. If two readers disagree, ask what each noticed. Perhaps the field is unclear, the source is mixed or the available context is insufficient. Revise the instruction where necessary and record the change.

Do not resolve every disagreement by choosing the more confident speaker. Ask for evidence from the source and the field definition. If the evidence remains insufficient, an uncertain value may be the most accurate outcome.

The immediate learning gain is clearer language about classification. Students practise explaining why a sample belongs in a group, instead of treating a label as an answer that needs no justification.

Contents · Previous chapter · Next chapter

Metadata can improve as a project develops. A learner may discover the original publication date, correct a mistaken source link or realise that two extracts came from the same document. The goal is not to freeze the first record forever. It is to make meaningful revisions traceable.

A small revision note can contain four elements: the sample ID, the field changed, the old and new values, and the reason. For example: “A01: publication year changed from unknown to 1995 after checking the original title page.” This note is far more useful than “fixed metadata.”

Separate a correction from a new interpretation

Changing a mistyped year after inspecting the source is different from redefining the category formal. The first repairs a recorded fact. The second changes the analysis rules and may require reviewing every sample classified under the old definition. Label the kind of change so that the next action is clear.

The TEI Header guidelines distinguish information about a resource, its source and encoding from its revision history. This documentation model shows why source description and later changes should both be visible. For classroom use, a dated note beside the record can carry the essential explanation.

Keep an unchanged copy of the original text when making a cleaned working version. If you remove a title, correct an obvious scanning error or standardise a spelling for a particular analysis, describe the operation. That makes it possible to revisit a surprising count and ask whether the difference came from language use or from preparation.

Do not burden the learner with a log of every cosmetic edit. Focus on changes that could affect selection, counts, interpretation or reuse. Renaming a decorative worksheet heading is different from removing a duplicated paragraph from the counted text.

At the end of the project, another person should be able to tell which version produced the reported figures. “Collection version two, with titles excluded and duplicate copies grouped” is a useful beginning. The Provenance Trail guide extends this habit across records and revisions.

Contents · Previous chapter · Next chapter

A detailed record does not make an unsuitable collection representative. You can document a narrow selection very carefully and still have a narrow selection. The benefit is that readers can see the limits, not that those limits disappear.

Suppose a class collects twenty restaurant reviews from one person’s blog. Recording dates, ratings, topics and lengths may support a thoughtful account of that writer’s language. It does not turn the material into a balanced sample of all restaurant reviews. A broad claim would require a collection designed for that broader question.

Comparable labels do not guarantee comparable material

Two groups labelled news could differ in topic, format, length, editorial practice or the number of independent sources. If one group consists mostly of short sports updates and the other of extended science reporting, a vocabulary difference might reflect those underlying differences.

Metadata can reveal this problem if relevant fields were recorded. It cannot establish which factor caused the difference merely by displaying several columns. A careful learner asks what has been held similar, what has changed and what remains entangled.

Nor does metadata settle the meaning of every occurrence. A source labelled science can include metaphor, quotation and everyday language. A word-frequency result still needs contextual reading. Use the Concordance Line guide when the next task is to inspect surrounding sentences.

Finally, metadata can contain mistakes. A date may have been copied incorrectly. A category may reflect an earlier rule. A source link may point to a later edition. Treat the record as documented information that can be checked, not as an authority that overrides the source.

A useful conclusion has two parts: what the collection shows under the stated rules, and what further evidence would be needed for a wider claim. For example: “These selected school notices often use direct action verbs. To discuss public notices more generally, we would need additional sources and a more systematic selection.”

This is a form of intellectual control. The learner can make a meaningful observation without claiming that a small project has resolved every question about English usage.

Contents · Previous chapter · Next chapter

A vocabulary project does not need private information to be interesting. Teacher-created passages, appropriately reusable public examples and carefully chosen published resources can support most introductory metadata lessons. Start with those before considering recordings or personal documents.

If a class uses its own work, explain the activity, follow the school’s applicable arrangements and keep access appropriate to the purpose. A classroom exercise about instruction verbs rarely needs students’ full names, addresses, family histories or private conversations. Neutral sample IDs are usually enough to connect records and texts.

A permission field is a reminder, not a substitute for permission

A cell saying “allowed” is unhelpful if nobody knows what was allowed, by whom or for which use. A short note should distinguish classroom reading from public redistribution. If the conditions are unclear, choose another source or check the relevant permission before sharing it more widely.

Likewise, a publicly readable webpage is not automatically material that may be copied wholesale into a downloadable collection. The practical lesson is to keep source and use conditions visible and to use appropriate material for the task. This guide is not a statement of legal permissions for particular texts.

When a learner quotes a short example in a report, include the source information needed to find and understand it. When summarising, distinguish the student’s own explanation from the original wording. When using constructed examples, say so explicitly. These habits preserve the relationship between evidence and presentation.

Do not ask students to infer sensitive attributes in order to fill a table. If the project is about sentence purpose, classify sentence purpose. If a research question would require detailed personal data, it belongs in a properly supervised setting with an appropriate design, rather than being improvised as an ordinary vocabulary activity.

The best beginner project leaves students with better questions, clearer writing and a small collection they can explain. Its value does not depend on collecting the largest possible quantity of information about the people behind the words.

Contents · Previous chapter · Next chapter

Different learners need different entry points. Some struggle to distinguish a sentence from a statement about its source. Others understand that distinction but overlook it during a comparison. More confident learners may be ready to design a collection and challenge its limits.

Repair: make the three layers visible

Use three cards: a teacher-created sentence, a source description and an interpretation. Ask the learner to explain which card tells us what was written, which tells us where it came from and which contains an analysis. Change one card at a time. Keep the sentence fixed while changing its possible purpose, then discuss what changes in the interpretation.

The checkpoint is an explanation, not fast sorting. A learner who says “This card describes the whole notice, not the meaning of this word” is showing the relevant distinction. If the terms are unfamiliar, begin with words, source note and our analysis before introducing the technical labels.

Stabilise: use a short repeatable routine

A suggested twenty-minute lesson can begin with three minutes identifying the question, five minutes checking source records, five minutes reading and counting examples, and five minutes writing a qualified conclusion. Use the remaining two minutes to explain one uncertainty. The timings are an adaptable teaching proposal, not a validated programme or a promise of results.

Give the learner a familiar mistake to repair: an upload year used as a publication year, an unknown audience treated as advanced or a summary treated as a transcript. Ask for the smallest change that makes the claim accurate. This encourages precise revision rather than abandoning the entire task.

Extend: ask what would change the conclusion

Let the learner propose an additional field or a better selection rule. They must explain how it would affect the investigation. Adding a field simply because it sounds technical is insufficient. A useful addition might distinguish independently written notices from copies of one template.

Across all three routes, finish with a real communication task: improve an instruction, explain a source comparison or write a short note to another reader. Metadata becomes useful when it supports that task. The learner does not need to memorise a large list of field names to begin using the underlying questions well.

Contents · Previous chapter · Next chapter

These exercises use invented situations. Work through them without looking ahead to the explanations. Write a short reason for each answer. Where the evidence is insufficient, state what information is missing rather than supplying a convenient guess.

Exercise 1: identify the layer

You see three statements: “Please close the lid”; “a teacher-created instruction”; and “close functions as a verb here.” Identify the text, metadata and linguistic analysis. Explain why the second statement does not establish the meaning of every word in the sentence.

Exercise 2: choose the relevant date

A letter was written in 1972, scanned in 2018 and added to a collection in 2026. A student is investigating language in letters written during the 1970s. Which date is most directly relevant? What additional question might still need checking?

Exercise 3: repair an overloaded field

A cell under the heading type contains “online children environment persuasive.” Propose separate fields that would make the record easier to interpret. Which part might need clarification rather than simple copying?

Exercise 4: count with the correct denominator

Two fictional groups contain the exact word finally. Group A has 12 occurrences in 3,000 tokens. Group B has 10 in 5,000 tokens. Calculate each rate per 1,000 tokens. Does the result by itself prove that Group A contains better writing?

Exercise 5: handle an unknown audience

Seven samples are marked beginners; three have an empty audience field. A learner labels the second group experts. Explain the error and suggest two reasonable ways forward.

Exercise 6: inspect independence

A set contains eight files. Four are copies of the same notice from different websites; four are independently written notices. How many files are present? How many distinct notice texts are known from this description? What should be recorded before claiming broad agreement across sources?

Exercise 7: choose the representation

You want to count how often a speaker used well in a recorded conversation. You have a transcript retaining repeated words and an edited summary. Which is the better starting point? What should you check about it?

Exercise 8: distinguish a correction from a rule change

One edit corrects a mistyped publication year. Another changes the project’s definition of instruction. Which edit may require rechecking the classification of many records, and why?

Exercise 9: write a bounded conclusion

In four teacher-created texts, direct action verbs appear more often in the two instructions than in the two stories. Write one accurate conclusion and one example of a claim that the evidence does not support.

Exercise 10: remove an unnecessary field

A class is studying sequencing words in published recipes. Its proposed record asks for each learner’s home address. Explain whether that information serves the question and suggest a more relevant field.

Contents · Previous chapter · Next chapter

Answer 1: distinguish the object from descriptions of it

“Please close the lid” is the text. “A teacher-created instruction” is metadata about its source and purpose. The verb label is a linguistic analysis of one occurrence. The source label does not settle every meaning because a whole-text classification and a word’s local function answer different questions.

Answer 2: attach the date to the event

The writing date, 1972, is most directly relevant to the stated question. The learner should still check whether the digital text preserves the original wording or contains later editing. Scanning and collection dates explain the digital record’s history, but they do not turn the letter into writing from 2018 or 2026.

Answer 3: split the dimensions

Possible fields are medium or publication setting, intended audience, topic and communicative purpose. “Online” needs clarification: it could mean originally composed for the web or merely available online later. “Children” also needs a project definition if a more specific audience distinction matters. Separate columns help only when their meanings are clear.

Answer 4: compare rates, then interpret purpose

Group A: 12 ÷ 3,000 × 1,000 = 4 occurrences per 1,000 tokens. Group B: 10 ÷ 5,000 × 1,000 = 2. Group A has twice the rate in this fictional set. That does not establish writing quality. Frequent finally might clarify a sequence, become repetitive or reflect the tasks represented. Read the examples before deciding.

Answer 5: missing is not the opposite category

An empty audience field does not mean experts. The class could check the sources and recover the audience information, or retain an explicit unknown group. It could also limit the report to the seven documented beginner samples. The important repair is to stop treating an absence of information as positive evidence for another label.

Answer 6: count files and distinct texts separately

There are eight files and five distinct notice texts under the stated description: one notice copied four times, plus four independent notices. Record the shared-source relationship and check the independence relevant to the claim. Four hosting websites do not automatically provide four independently authored endorsements of the wording.

Answer 7: inspect the transcript’s preparation

The transcript retaining repeated words is the better starting point. Check its completeness and accuracy against the recording where available, and confirm what its transcription conventions retain or omit. The summary is evidence about how someone represented the conversation, not a dependable count of the speaker’s exact words.

Answer 8: a new definition can change membership

Changing the definition of instruction may require reviewing every record whose classification depends on it. Correcting one mistyped year normally concerns that record, although it may also change its date-group membership. Both edits should be traceable, but their consequences differ.

Answer 9: describe the set you actually analysed

An accurate statement is: “In our four constructed texts, the instruction group contains more direct action verbs than the story group.” An unsupported statement is: “All good instructions must contain more action verbs than every story.” The classroom set demonstrates a comparison method; it does not establish a universal writing rule.

Answer 10: collect information about the relevant source

Learners’ home addresses do not help answer the sequencing-word question. Remove that field. A more relevant field might identify the recipe source, intended readership, length or whether the extract includes the full procedure. Good metadata serves a purpose; it is not a competition to collect personal detail.

Contents · Previous chapter · Next chapter

Choose one small question that leads to a writing or reading decision. For example: “How do instructions tell the reader what to do next?” Select a few appropriate texts, record their sources and purposes, and identify the expressions that organise actions. Keep the project small enough that every example can be read.

First session: build and explain the collection

Agree on the question and the inclusion rule. Create a record for each sample. Read the material before counting anything, and note whether the sample is complete, adapted or copied. The first checkpoint is whether another learner can explain why each text belongs in the collection.

Second session: investigate and revise a real passage

Inspect the target expressions in context. Compare their jobs rather than simply collecting totals. Then revise a short instruction written by the learner. Ask what changed and why. Perhaps a vague then becomes “After the glue has dried,” because the action depends on a condition rather than on time alone.

Return later with a different source

On another day, introduce a new example with a deliberate complication: a story containing instructions, a revised edition or a source with an unknown date. Ask the learner to identify what additional information matters. This tests whether the method transfers beyond the original worksheet.

Keep a short progress note focused on observable behaviour. Can the learner distinguish source notes from interpretation? Can they explain a date? Can they calculate a simple rate correctly? Can they qualify a conclusion without making it empty? Can they use the evidence to improve a piece of communication?

Difficulty in one area suggests a specific next step. A calculation error needs numerical practice. A mixed-up date needs source-reading practice. A broad conclusion needs work on scope. Calling all three “weak vocabulary” would hide the repair.

The project succeeds when the learner can use words such as source, sample, context, unknown and revision to make a task clearer. The record supports thinking; it should not become a barrier that makes a modest reading question feel unnecessarily technical.

Contents · Previous chapter · Next chapter

Is corpus metadata the same as a bibliography?

A bibliography identifies sources. Corpus metadata can include bibliographic information, but it may also describe text purpose, medium, sample extent, processing and collection decisions. For a small exercise, a source citation plus a few well-chosen fields may be enough. The question determines the necessary detail.

Does every text need exactly the same fields?

A shared core makes comparison easier, but some fields apply only to certain material. Recording date is relevant to a recorded conversation; edition information may matter for a revised book. Define how the project distinguishes not applicable from unknown so that empty cells do not create misleading groups.

Can a document have more than one genre or purpose?

Yes. A webpage might combine an explanation, instructions and a personal account. A project can label sections separately, permit multiple values or identify a dominant purpose. The important point is to state the rule and avoid treating a mixed document as uniformly one kind of language.

Can I infer metadata from the wording?

Sometimes the text provides evidence for a classification, such as the presence of numbered instructions. Other details, including authorship or personal background, may not be recoverable from wording alone. Distinguish observed information, reasoned classification and uncertainty. Do not give a guess the appearance of an established source fact.

Is more metadata always better?

No. Irrelevant fields consume time and may collect information that the task does not need. Begin with fields that help select, interpret, verify or responsibly reuse the material. Add detail when a concrete question makes it useful, and explain the field so that it can be filled consistently.

Can metadata prove that a word is appropriate?

It can show that an example comes from a situation relevant to your task. You still need to read the sentence, understand the word’s function and consider your own audience. A source category cannot make every expression in a text suitable for every new situation.

What if a useful source has incomplete information?

Record the gap and decide whether it affects your question. An unknown publication date may prevent a historical comparison while leaving the text useful for studying sentence structure. Check the source where practical, keep an unknown category or narrow the claim. Avoid inventing a value to complete the table.

Do I need XML or specialist software?

No. A few texts and a clearly labelled table can teach the core ideas. Formal encoding systems support larger projects and exchange between tools. Learn the distinctions first. Introduce technical formats when they solve a real problem, such as managing many linked samples consistently.

Can constructed examples count as a corpus?

You can organise constructed examples into a collection for teaching or testing a method. Their status must remain explicit. They show how a procedure works, but they do not independently demonstrate how a population of speakers or writers normally uses language. A claim about observed usage requires appropriate observed material.

What is the most useful beginner habit?

Before using an example as evidence, ask: “Where did these words come from, and what question can that source help me answer?” Then record the answer briefly enough that someone else can follow it. This habit is useful in vocabulary learning, reading comprehension and research across subjects.

Contents · Previous chapter · Next chapter

The sources below document real approaches to describing language resources. The classroom records, numerical datasets, exercises and teaching routines in this article are original constructed illustrations. Their figures are not reported corpus findings, and the suggested fields are not presented as a mandatory research standard.

• Text Encoding Initiative: The TEI Header documents descriptions of resources, sources, encoding and revisions.
• Text Encoding Initiative: Language Corpora discusses corpus design and contextual information at collection and component levels.
• British National Corpus User Reference Guide: The Header shows how a particular corpus records information about its texts and the collection.
• About the British National Corpus describes the original corpus’s scope, material and sampling.
• American National Corpus: First Release Encoding Conventions provides a project-specific example of headers and separate linguistic annotations.
• Language Bank of Finland: The Korp Corpus Input Format describes declared token attributes and structural metadata in a technical corpus format.

For the broader subject, continue with Corpus Linguistics. For a new learning route, use the Vocabulary Articles and Knowledge Hub or the Vocabulary Learning Hub.

Take one sentence you are considering as a model for your own writing. Write down its source, purpose and any uncertainty that affects your decision. Then explain what you can learn from it. That short record turns an isolated example into evidence you can use with greater control.

Contents · Previous chapter · Sources and further reading

Sources and further reading

These sources support the definitions and technical distinctions. The worked examples, classroom datasets and teaching activities are original unless explicitly identified otherwise.

• TEI Guidelines: The TEI Header — Primary standard documentation; source/encoding/profile/revision distinctions.
• TEI Guidelines: Language Corpora — Primary standard documentation; design criteria and collection/component descriptions.
• BNC User Reference Guide: The Header — Corpus producer documentation; real header implementation.
• About the British National Corpus — Corpus producer description; scope, sampling and contextual information.
• American National Corpus: First Release Encoding Conventions — Corpus producer documentation; metadata and linguistic annotations.
• Language Bank of Finland: The Korp Corpus Input Format — Corpus service documentation; token versus structural attributes.

Back to contents · Vocabulary Articles and Knowledge Hub