A concordance sample is a selected set of occurrences from a larger corpus search result. Each concordance line shows a word or phrase with its surrounding text. The sample makes thousands of possible examples manageable for human inspection, but the selection method—first results, random selection, document-balanced or purposefully chosen examples—determines what conclusions the learner can honestly draw.
Did you know the first twenty results can be misleading even when every sentence is genuine? A corpus may arrange lines in source order, placing many examples from one writer or topic together. A student who mistakes that page for typical English may learn a narrow rule and apply it too widely. Sampling makes vocabulary a technology for testing evidence: it teaches how to select observations, separate word senses and qualify a conclusion before using a phrase in real writing.
This guide explains what concordance sampling is, why shuffling differs from selecting a random subset, how genre and document duplication affect conclusions, and when deliberate teaching examples are better than estimated frequencies. It includes original classroom exercises for Primary and Secondary English, Science, Mathematics, History, General Paper, translation and professional communication. The first three paragraphs describe the method; the detailed research notes and authoritative resources appear below.
Evidence note: Any classroom dataset, count or percentage presented as an example here is invented and labelled for teaching. The guide does not claim to have measured an English-language population or to have executed the illustrated query.
Contents: how to choose word evidence without fooling yourself
Open 44 chapters and practice checks
- 1. Why a concordance sometimes shows too much
- 2. The first results are not necessarily typical
- 3. Sampling is a decision made after retrieval
- 4. Random sampling has a specific purpose
- 5. Shuffling and sampling are different
- 6. The denominator is the full match set
- 7. A sample can still contain wrong matches
- 8. Sample size depends on the question
- 9. The first twenty lines problem
- 10. Convenience sampling is quick but narrow
- 11. Purposeful examples are not dishonest
- 12. Stratified sampling protects categories of interest
- 13. Document-balanced samples ask a different question
- 14. Duplicates inflate visible agreement
- 15. A source label can prevent a false generalisation
- 16. Sample across senses only when that is the goal
- 17. A negative occurrence is not always a new sense
- 18. Quoted language can dominate a search
- 19. Context windows can create classification errors
- 20. Sample the units you intend to describe
- 21. Random ordering is not a guarantee of independence
- 22. A rare sense may disappear from small samples
- 23. Sample confidence is not automatic
- 24. Annotation disagreement affects sample analysis
- 25. Blind review can reduce expectation effects
- 26. Use a query log before changing selection rules
- 27. The difference between coverage and illustration
- 28. A class exercise using invented cards
- 29. Primary English: familiar word with several senses
- 30. Secondary English: strengthen analytical verbs
- 31. Science: distinguish observation from inference
- 32. Mathematics: identify true logical conditions
- 33. History: source attribution belongs in the sample
- 34. General Paper: counterarguments cannot be ignored
- 35. Translation: one phrase may have several functions
- 36. Workplace documents: approval terms have consequences
- 37. Singapore English needs suitable source coverage
- 38. A source might have licensing restrictions
- 39. Privacy matters in student corpora
- 40. A sample audit table keeps evidence honest
- 41. A practical fifteen-minute investigation
- 42. When to ask for a second sample
- 43. What not to conclude from sampled lines
- 44. The closing test: select, inspect, transfer
One fictional search, three different samples
| Selection rule | What the teacher sees | What can be claimed |
|---|---|---|
| First 20 concordance lines | Whatever occurs at the beginning of the stored search results | These 20 lines illustrate usages in the displayed portion |
| Randomly selected 20 lines | A probability-based subset under the tool’s sampling procedure | Observed patterns in the sample; stronger inference requires assumptions and uncertainty analysis |
| Deliberately selected 4 senses | One useful example of each intended meaning, if available | The selected senses can be illustrated; their real frequency is not measured |
All numbers above describe a fictional lesson. The key distinction is between illustrating possibilities and estimating their relative prevalence. Neither goal is wrong; hiding which goal was used is the error.
1. Why a concordance sometimes shows too much
A concordance is a set of occurrences of a searched expression with neighbouring words shown. If a corpus contains tens of thousands of matches for run, a learner cannot reasonably read every line in a short lesson. A concordance sample is a selected subset of those occurrences, chosen to make analysis manageable. The selection procedure shapes what the learner sees.
The important distinction is between the population of matches returned by the query and the sample of lines actually inspected. A sample may be useful for discovering word senses and grammatical patterns, but one screen of examples cannot be silently described as every use of the word in English.
2. The first results are not necessarily typical
Did you know that a concordance interface may initially show lines in the order they occur in the corpus? Sketch Engine documents this default and provides shuffle and random sample controls. The earliest matches might all come from one document, author or topic section.
A learner who reads just the first twenty lines could mistake one writer’s repetition for a widespread usage pattern. This is not evidence that the software is dishonest: the interface is displaying a particular order. The researcher must decide whether that order matches the question.
3. Sampling is a decision made after retrieval
The corpus was assembled from texts; a search query retrieved matching occurrences; and a concordance sample selects some of those matches for inspection. These are three separate stages. Confusing them makes it difficult to describe where evidence came from.
A random sample of matches does not repair a corpus made entirely of one genre. Likewise, a carefully assembled corpus does not ensure that the first displayed lines are representative of all retrieved occurrences. Good vocabulary research keeps selection at each stage visible.
4. Random sampling has a specific purpose
A random sample attempts to give the eligible occurrences a specified chance of selection under a defined procedure. When properly implemented and used, random selection can reduce the tendency to favour lines near the beginning or those that look memorable.
Randomness does not guarantee that every sense, register or genre will appear in a small set. A rare meaning might be absent merely by chance. The learner should record the sampled number of lines and avoid treating nonappearance as proof that the rare sense is impossible.
5. Shuffling and sampling are different
Shuffling rearranges the order of a result set. Sampling selects a subset. In Sketch Engine, both options are provided and described separately. A shuffled screen can make it easier to browse a wider range of contexts, but the display size and inclusion rules still matter.
A student should ask: Did the tool just reorder existing lines, or did it reduce them to a smaller set? If a classmate tries to reproduce the result, those operations could produce different visible sequences. Precise vocabulary includes precise names for actions.
6. The denominator is the full match set
Suppose an invented query returns 1,000 candidate occurrences and the researcher inspects a hundred. Those hundred are the sample; the full thousand are the retrieved set. A student should not report a sense percentage from a convenience sample as though it were a census.
A useful note states both sizes and selection method. If 15 of the inspected hundred show a particular use, the statement supported without more is that 15 sampled examples were classified that way. Generalising to all thousand requires an appropriate sampling method and attention to uncertainty.
7. A sample can still contain wrong matches
An exact spelling such as bank can refer to money or a riverbank. The search may retrieve both. Selecting a random subset does not automatically separate the intended meanings, because randomness concerns selection, not semantic accuracy.
Classify each occurrence by sense after opening enough context. Keep ambiguous cases visible rather than forcing them into a preferred category. A transparent category such as ‘insufficient context’ is often more reliable than guessing.
8. Sample size depends on the question
A small classroom exercise may need five well-chosen examples to demonstrate that a word has more than one sense. A research claim about relative frequencies or genre differences needs a stronger design and typically more evidence.
The number itself is not a quality certificate. Five examples can teach a contrast without supporting a population estimate. Hundreds of poorly selected lines can mislead if they all come from the same copied source. Match the sample to the purpose and claims.
9. The first twenty lines problem
Imagine an invented school-language corpus organised by date. The first twenty matches for charge come from an electricity lesson. A learner might infer that the word chiefly means supplying electrical energy. Later matches from finance and legal texts could contradict that impression.
If the objective is to teach several senses, sample across relevant source types or deliberately select illustrative examples while labelling the choice. If the objective is empirical frequency, use an appropriate probability sampling strategy. A classroom demonstration and an unbiased prevalence estimate are different projects.
10. Convenience sampling is quick but narrow
A convenience sample consists of examples chosen because they are readily available: the first page of results, a favourite textbook or the most understandable lines. Such examples can be useful when the teaching objective is simply to explain a construction.
The limitation is selection bias. Easy examples may omit difficult cases that matter to real reading. A responsible educational article labels a convenience sample rather than presenting it as a balanced picture of English.
11. Purposeful examples are not dishonest
A teacher may choose one example of record as a noun and one as a verb because that contrast is the lesson. This deliberate sampling is pedagogically sensible. It does not claim that the two usages occur equally often in a corpus.
Say that the sentences were chosen to illustrate distinct roles. Students should learn that evidence for a definition, evidence of a possibility and evidence of a frequency pattern require different methods. The difficulty begins when an illustrative selection is misrepresented as a measured distribution.
12. Stratified sampling protects categories of interest
A stratified design divides eligible occurrences into relevant groups, such as genre or decade, and samples within each group. That can ensure analysts inspect more than one kind of text, even when a large category dominates the full set.
The method must be reported. If equal numbers are taken from small and large genres, the sample’s raw proportions do not represent the corpus’s original composition. A weighted estimate may be necessary for certain population claims. Equal representation for teaching and proportional representation for estimation are different aims.
13. Document-balanced samples ask a different question
If one novel contains two hundred matches and twenty shorter texts contain one each, a sample of occurrences can be dominated by the novel. A document-balanced selection might limit how many examples are taken from any one source.
That can be useful for inspecting a diversity of authors or contexts, but it changes what the sample represents. If the research question concerns the proportion of all tokens used in a corpus, document-balancing without appropriate adjustments may distort the target. Define the question first.
14. Duplicates inflate visible agreement
Syndicated news, repeated website templates or copied exercises can produce dozens of nearly identical concordance lines. A random sample of occurrences can still include those duplicates because they genuinely occupy many positions in the stored corpus.
Check whether the aim is to describe stored texts or independent language use. Where needed, identify duplicates and near-duplicates before sampling. Do not count copied lines as independent confirmations of one linguistic claim.
15. A source label can prevent a false generalisation
Corpus metadata can identify genre, document, date or other source features. Inspecting those labels alongside sampled lines helps readers see whether the evidence is drawn from letters, news, school reports or fiction.
A learner may discover that a phrase is characteristic of one subset, which is a useful finding. The conclusion should name that subset. ‘We saw this in the sampled school notices’ is more responsible than ‘everyone uses it this way’.
16. Sample across senses only when that is the goal
For a dictionary-style lesson on light, the teacher may deliberately include a noun, adjective and verb example. This communicates the range of roles. It should not be interpreted as a probability distribution for those uses.
If the aim is to estimate which sense is most frequent in a corpus, the sampling strategy must avoid deliberately overrepresenting minority senses. One cannot use the same chosen examples both to guarantee diversity and to calculate unbiased prevalence without adjustment.
17. A negative occurrence is not always a new sense
A sentence such as ‘The evidence did not support the claim’ contains the expression support the claim, yet it denies the proposition. When classifying concordances, mark negation and reported speech instead of treating the line as positive evidence for the claim.
This is essential in Science, History and public communication. The search retrieved a phrase, not the writer’s endorsement. A concordance sample becomes meaningful only after the full clause is read.
18. Quoted language can dominate a search
A long article may quote a controversial statement repeatedly while criticising it. A search for the statement’s key word can return many instances, but those occurrences may not represent the author’s own vocabulary preferences.
Annotate speaker and quotation status for consequential interpretations. A sample containing only excerpts could misrepresent the surrounding argument. Attribution is a necessary part of vocabulary evidence.
19. Context windows can create classification errors
A KWIC line shows limited text around the target. The next sentence might clarify the intended meaning, identify the actual speaker or explain a technical condition. Classifying the sample from tiny fragments can introduce errors.
Expand unclear examples to sentences or paragraphs before judging them. If the full context is unavailable, record that limit. A larger sample of poorly understood fragments is not automatically stronger than a smaller set of carefully interpreted examples.
20. Sample the units you intend to describe
A ‘sample of sentences’ differs from a ‘sample of occurrences’. One sentence might contain the target word three times and therefore generate multiple concordance hits. A ‘sample of documents’ is different again.
State whether the unit is a token occurrence, a sentence, a document or an author. Mixing these units can exaggerate frequencies and make an apparent variety of examples less independent than it looks.
21. Random ordering is not a guarantee of independence
Two randomly selected lines can come from the same document or repeated source. Randomness helps with ordering or selection, but it does not necessarily make observations independent.
If the research question concerns author-level generalisations, record author or document clusters. A classroom does not need advanced statistical models for every lesson, but should avoid claiming that a hundred hits represent a hundred separate people.
22. A rare sense may disappear from small samples
A corpus might contain a rare technical meaning alongside a dominant everyday meaning. A small random sample can miss that technical sense even when it is present. This matters when a student is looking for a precise specialist use.
For teaching domain vocabulary, use a domain-specific query or deliberately locate authentic specialist examples. Label the selection as targeted. For prevalence claims, explain the limits of a sample that may overlook rare categories.
23. Sample confidence is not automatic
Researchers sometimes calculate uncertainty intervals or use statistical models for sample-based estimates. Those results depend on the selection procedure, sample size, independence and assumptions. A convenience sample does not become probability evidence merely because someone attached a percentage.
A school article should explain uncertainty in plain language before introducing formal calculations. ‘This pattern appeared in our sample’ is often sufficient. Only claim population percentages or confidence levels when the actual design supports them.
24. Annotation disagreement affects sample analysis
Two readers may assign different meanings or registers to the same line. Rather than averaging guesses, they can write clear category definitions, review ambiguous examples and record disagreements.
For research, inter-annotator agreement can be measured under a defined method. For teaching, discussion itself may reveal a missing context or an imprecise category. The sample must be interpreted consistently as well as selected responsibly.
25. Blind review can reduce expectation effects
If a student expects a word to be negative, they may notice examples confirming that view. Concealing the intended hypothesis from an independent annotator, where practical, can help test whether the categorisation is being driven by expectation.
Blind review is not always feasible in classrooms. A simpler safeguard is to require one counterexample and an explicit explanation of ambiguous lines. Honest vocabulary learning welcomes inconvenient evidence.
26. Use a query log before changing selection rules
A researcher might discover a pattern, narrow the query, choose a more dramatic sample and then forget that the method changed. Keeping a log preserves the full initial query, search settings, sample rule and later revisions.
This protects against presenting an exploratory discovery as a prediction made in advance. A clear analysis can change direction; it just needs to show where the changes occurred. Reproducibility is an intellectual habit, not merely a software feature.
27. The difference between coverage and illustration
A student might need an illustration of a collocation such as strong evidence. One or two credible examples can serve. To claim that a collocation occurs in many kinds of English, the student needs wider corpus coverage and a defined comparison.
The lesson objective should appear at the beginning. If an example merely shows that a phrase is possible, do not convert it into ‘most people say this’. The difference separates teaching illustration from linguistic generalisation.
28. A class exercise using invented cards
Write twelve original sentences containing claim as a verb, noun and part of a quotation. Place each on a card and number the cards. Shuffle the cards, select a subset and ask pupils to classify the meaning and speaker.
Repeat the activity with another subset. Discuss what changed and what remained uncertain. Because the cards are invented, the result is a demonstration of sampling variation rather than a measurement of English usage. That explicit distinction is part of the lesson.
29. Primary English: familiar word with several senses
Use bright in examples about light, colour and a person’s quick understanding. Choose sentences to illustrate each sense, then ask children to identify contextual clues.
Do not claim that the teaching cards show which sense is most common. For a learner whose challenge is comprehension, varied examples may be more valuable than prevalence estimates. Accuracy of sense selection is the success criterion.
30. Secondary English: strengthen analytical verbs
A student may overuse shows in literary analysis. A sample of authentic, relevant analytical writing can help explore alternatives such as suggests, reveals and contrasts.
Every proposed replacement changes the strength or nature of the claim. The tutor should ask what the passage actually supports and whether the new verb preserves it. A sample improves writing only when evidence and meaning remain aligned.
31. Science: distinguish observation from inference
A Science corpus may contain many examples of indicates or demonstrates. Selecting examples from published explanations can teach how authors position evidence, but cannot prove the student’s own experiment succeeded.
Use the actual experiment, measured values and uncertainties to decide whether the chosen verb is defensible. A corpus sample supplies language patterns; the scientific method supplies evidential justification.
32. Mathematics: identify true logical conditions
In a class exercise, sample questions containing at least, at most and more than. Students classify how each changes the set of permitted values.
Do not treat the sample as proof of official examination prevalence. The teaching objective is to convert words into inequalities accurately. The final test is an unseen problem solved with the correct boundary.
33. History: source attribution belongs in the sample
A historical text may use the word reform to quote a policy’s supporters or to challenge their description. A line without its source or quotation status can be misleading.
When selecting examples, preserve dates, authors, audiences and a sufficient passage. Students should compare viewpoints while distinguishing historical evidence from the vocabulary used to frame it. A balanced selection alone does not guarantee balanced interpretation.
34. General Paper: counterarguments cannot be ignored
A GP essay benefits from accurate attribution and calibrated claims. A corpus sample that contains only supportive statements may conceal the debate around a term or policy.
Include relevant contrary instances if the aim is to investigate framing or argumentation. That does not mean choosing an equal number from every viewpoint in a prevalence study; it means making the sampling objective and evidence boundaries explicit.
35. Translation: one phrase may have several functions
A polite request may be translated differently in a formal letter and a friendly message. Sampling only a small set of casual dialogues can lead a translator to choose an inappropriate target form for an official notice.
Sample relevant target-language contexts and check the source meaning first. The translation should preserve speaker, request, obligation and tone. Vocabulary evidence is helpful when it is comparable to the communicative task.
36. Workplace documents: approval terms have consequences
The words reviewed, accepted and approved may mark different actions. A corpus sample of old messages can suggest sentence patterns but cannot establish that a current approval occurred.
Check the operative policy and actual record. A well-chosen example helps with drafting; it must never replace the evidence of authorisation. This is where precise vocabulary becomes operational safety.
37. Singapore English needs suitable source coverage
Different varieties of English have legitimate patterns, and a British or American corpus may not include all Singapore contexts. A sample from another region should not be used to declare a local expression defective.
For examination writing, teach the required register while acknowledging how Singapore English functions in daily interaction. The sample should match the teaching goal, and any claim about a community should be supported by relevant evidence.
38. A source might have licensing restrictions
A concordance line can contain copyrighted material. A corpus service’s licence may restrict how many examples may be published or reproduced. A teacher should understand the rights attached to the source texts.
In a public article, write original illustrative sentences when possible and link to the original corpus resource for further investigation. Do not present constructed examples as if they were retrieved lines. Provenance and permissions are part of responsible vocabulary research.
39. Privacy matters in student corpora
A school collection of genuine compositions might include names, addresses or personal incidents. Sampling does not remove those identifiers, and even a short concordance line can expose confidential information.
Use public, licensed corpora or invented examples for general demonstrations. For authorised classroom research, follow consent, de-identification and storage policies. Educational value does not override students’ privacy.
40. A sample audit table keeps evidence honest
For each selected example, record an ID, original document, genre, date where known, relevant sentence, sense classification and any uncertainty. Add whether the line was selected randomly or deliberately.
This table makes a small sample reviewable. Another person can ask whether the cases cover the intended contexts or whether one source dominates. A neat summary without such a trail may be persuasive but impossible to verify.
41. A practical fifteen-minute investigation
Spend a few minutes naming the precise word-choice problem. Retrieve examples from a relevant source and choose a manageable sample by a stated rule. Classify senses, inspect confusing cases and revise one real sentence.
End by asking the learner to write another original sentence without copying the example. The timing is an illustrative teaching plan, not a validated claim about learning outcomes. Transfer is the real measure of success.
42. When to ask for a second sample
If the first sample appears dominated by one genre, contains unexplained contradictions or fails to include expected senses, a second sample or adjusted design may help.
Record why the change was made. Do not conceal the initial observations or present the final selection as a single unbiased draw. An honest refinement can be a valuable part of the lesson.
43. What not to conclude from sampled lines
A sample cannot alone tell us whether a statement is true, whether the corpus represents every speaker, whether an author’s motivations are known or whether a phrase is universally grammatical.
Those are separate questions requiring source criticism, grammatical analysis and, where appropriate, stronger evidence. Vocabulary research is most credible when it does not ask a small set of examples to prove more than they can.
44. The closing test: select, inspect, transfer
After studying a concordance sample, the learner should explain how the lines were chosen, name one limitation and use the intended expression appropriately in a fresh sentence.
That is vocabulary functioning as a practical technology. It helps people move from a guess to evidence, from evidence to a qualified conclusion and from that conclusion to better writing. A sample is a tool for seeing clearly, not a licence to overstate what has been seen.
A reproducible concordance sampling worksheet
- Name the intended word sense, grammatical pattern or teaching question.
- Record the identified corpus and the exact query, including lemma or surface-form settings.
- Count the full retrieved occurrences before sampling where the tool reports the number.
- Choose a sample rule: random selection, a documented stratification or purposeful examples for teaching.
- Record sample size and source IDs; inspect duplication and concentration in one document.
- Read complete sentences, marking negation, quotation, speaker and uncertain meanings.
- State conclusions at the level justified by the actual selection rule.
- Create an original transfer sentence and explain why the selected expression fits the real reader.
Frequently asked questions
1. What is a concordance sample?
A selected subset of concordance matches chosen for detailed analysis or instruction.
2. Why not read just the first search page?
Its order can reflect the corpus or interface rather than a representative variety of uses.
3. Does random sampling guarantee every meaning appears?
No. Rare meanings may be missed, especially in small samples.
4. Can I choose examples deliberately for teaching?
Yes, if you label them illustrative and do not infer population proportions from them.
5. What is the difference between shuffle and sample?
Shuffling changes order; sampling selects a subset under a stated rule.
6. Can repeated lines bias a sample?
Yes. Duplicated documents and templates can overrepresent particular expressions.
7. Is a corpus sample the same as the corpus sampling frame?
No. One selects retrieved occurrences; the other concerns which original texts could enter the corpus.
8. Why expand KWIC context?
Short fragments can hide negation, quotation or a different sense.
9. Can a sample prove that a claim is true?
No. It records how expressions occur, not whether the propositions they convey are correct.
10. What is the practical learning outcome?
An accurately chosen phrase, a clear explanation and recognition of the evidence limits.
Sources, limitations and further study
- Sketch Engine: Concordance Guide — official explanation of result order, shuffle and random sampling features.
- Natural Language Toolkit: Concordance Usage — examples of matching words and inspecting contexts.
- AntConc Manual: Concordance — KWIC search, context and sorting.
- Sketch Engine: Word Sketch — grammatical partner summaries that complement line-level inspection.
- Vocabulary | What Is a Corpus Sampling Frame? — where the original texts come from.
- Vocabulary | What Is a Corpus Sampling Weight? — how sampling proportions affect certain estimates.
- Vocabulary | What Is a Concordance Line? — read the evidence behind a match.
- Vocabulary Learning Hub — the continuing school-to-life vocabulary pathway.
A concordance sample is valuable when it helps us see and evaluate real word use. A responsible reader knows what was selected, what was left unseen and why the evidence supports a particular word choice. That understanding travels beyond vocabulary into media literacy, research and everyday decisions.