A corpus query is a precisely stated search for a word, phrase or grammatical pattern inside a defined collection of texts. The query can request an exact spelling, inflected forms grouped under a lemma, or a sequence of words with selected properties. Its results are examples to inspect, not automatic proof of a universal usage rule.
Why does this matter? Students may know a dictionary definition and still misunderstand the word inside a passage or choose an unnatural partner in a composition. A corpus query turns ‘Does this expression work here?’ into a question that can be investigated with language evidence. It helps learners select meanings, understand grammar and write for a real audience without treating vocabulary as a list of impressive synonyms.
This guide explains the what, why and how of corpus queries: exact forms, lemmas, phrase searches, Corpus Query Language, filters, concordances, false positives, false negatives, source selection, interpretation and practical classroom use. It includes original exercises for Primary and Secondary English, Mathematics, Science, History, translation and adult communication. Official corpus-tool documentation is linked below.
Editorial note: Classroom cases and sentences in this article are original illustrations. No fabricated search result or frequency is attributed to a real corpus.
Contents: from a word to a testable question
Open 39 learning chapters
- 1. A corpus query is a question translated into a search
- 2. Why vocabulary is more than a list of definitions
- 3. Define the source before searching
- 4. A research question is narrower than a keyword
- 5. Search for an exact word form
- 6. Search for a lemma
- 7. Search for an ordered phrase
- 8. Word proximity does not always mean grammar
- 9. Regular expressions describe character patterns
- 10. Corpus Query Language is more structured
- 11. How a CQL sequence works
- 12. Grammar tags narrow the search
- 13. Metadata can make a subcorpus
- 14. Context windows can mislead
- 15. Sorting a concordance uncovers candidates
- 16. Filtering can become cherry-picking
- 17. Counts depend on the query definition
- 18. No match does not mean impossible English
- 19. False positives require sense checking
- 20. False negatives require alternative forms
- 21. Precision and recall are different
- 22. Test simple conditions before complicated ones
- 23. Case sensitivity is a design choice
- 24. The corpus itself can contain errors
- 25. Primary English: word senses on paper
- 26. Secondary English: improve a composition
- 27. Mathematics: operational vocabulary
- 28. Science: keep evidence and inference apart
- 29. History: words can carry a viewpoint
- 30. Translation: preserve the communicative act
- 31. Policy and workplace vocabulary
- 32. Keep student data private
- 33. AI can suggest searches but cannot claim they ran
- 34. Exploration and formal testing differ
- 35. A useful query notebook
- 36. A fifteen-minute learning cycle
- 37. When a result is unexpectedly strange
- 38. Do not confuse frequency with correctness
- 39. The final test is a better sentence
Four illustrative kinds of query
| Search | Example in compatible CQL | Learning purpose |
|---|---|---|
| Word form | [word="went"] | Target one visible form |
| Lemma | [lemma="go"] | Target relevant inflected forms where supported |
| Phrase | [lemma="refill"] [lemma="the"] [lemma="teapot"] | Target an ordered three-token sequence |
| Grammar-aware | A documented part-of-speech filter | Exclude irrelevant noun or verb uses |
These examples reproduce the *kind* of syntax documented by Sketch Engine. They were not executed for this article, and the available attributes and tagsets vary among corpora.
1. A corpus query is a question translated into a search
A corpus query is an instruction for finding occurrences of words, phrases or grammatical patterns inside a defined text collection. The human question comes first. For example, a student might ask how the word evidence is used in school explanations rather than in ordinary conversation. The query is the operational version of that question, not its answer.
Once the query returns examples, the learner still has to read them, classify senses, identify their sources and decide whether they apply to the intended writing task. This distinction matters: retrieving a sentence is an observation; claiming that it represents correct or typical usage requires further reasoning.
2. Why vocabulary is more than a list of definitions
A dictionary tells us what a word can mean, but it cannot by itself guarantee that the learner will use the word in the right grammatical relationship. A student might know the noun decision and still pair it with an unsuitable verb because the learner has not practised the actions that can be performed with decisions.
A corpus query gives the learner a repeatable way to inspect word behaviour. Instead of accepting ‘that sounds wrong’, teacher and pupil can compare traceable examples. The outcome should be a clearer sentence and a rule the learner can explain, not an impressive display of software settings.
3. Define the source before searching
A corpus can contain newspaper writing, fiction, school essays, transcribed conversations or specialist documents. Every collection has a boundary: it includes some speakers and situations but excludes others. A query run over a corpus of legal documents does not describe every possible way Singapore pupils use English.
Record the corpus name, approximate date range, variety of English, genre and any relevant sampling information. Those details affect the meaning of a result. A frequent construction in technical reports may be unhelpful when a child is trying to write dialogue for an everyday school story.
4. A research question is narrower than a keyword
Entering charge into a search box is not enough to decide whether the learner is studying a price, a battery, a criminal accusation or rapid movement. The written form is shared, but the senses and grammatical frames differ. A collection of mixed matches can obscure the pattern.
Write a precise goal such as ‘I want examples of charge meaning to request payment.’ Then classify the retrieved sentences. The better question will reduce false matches and prevent the teacher from promoting a combination that belongs to the wrong meaning.
5. Search for an exact word form
An exact surface-form query targets a visible sequence of letters, such as went. It is appropriate when a student needs a particular tense or form rather than every member of the word family. Some systems also offer case-sensitive matching, while others ignore capitalisation by default.
Exact matching is a precision tool, but it can miss useful alternatives. A learner searching for write may not see occurrences of wrote or written. The notebook should record whether the query matched only the surface spelling or a broader family.
6. Search for a lemma
A lemma is the base entry under which inflected forms are conventionally grouped. With suitable annotation, a lemma query for go can retrieve goes, going, went and gone. This is useful when the question concerns a verb’s general behaviour.
However, lemmatisation depends on the software’s analysis and may contain mistakes. If the purpose is to study a particular tense, a broad lemma search might include material irrelevant to the lesson. Choose the tool setting that matches the question, then verify a few actual examples.
7. Search for an ordered phrase
A phrase query requires a sequence such as as a result or take into account. Those expressions can perform functions that the individual words do not explain separately. Phrase searching is especially helpful for discourse markers and recurring academic frames.
The limitations matter. An exact phrase query may omit a legitimate variant that contains an additional word, punctuation or changed inflection. If the learner wants all possible ways to express the function, compare documented alternatives rather than declaring one fixed sequence the only correct version.
8. Word proximity does not always mean grammar
Two words can appear beside each other by coincidence. A sentence might contain an adjective next to a noun without the adjective describing that noun, especially across punctuation, quotations or unusual constructions. A simple window query can return these instances.
Before teaching a collocation, check the actual relationship. Read the complete phrase and, where useful, a grammatical dictionary entry. If a program classifies subjects and objects automatically, treat its output as a hypothesis to inspect rather than a perfect description of what each sentence means.
9. Regular expressions describe character patterns
A regular expression allows a search system to match a family of character sequences. It may help locate alternate spellings or forms with a common visible prefix. The meaning of special symbols depends on the regex engine and options; copying patterns blindly between tools can produce surprising results.
A broad prefix query can retrieve proper names, unrelated terms or forms the student did not intend. Test the expression against a few known positive and negative examples before using its counts. A search that retrieves more matches is not necessarily a search that answers the question better.
10. Corpus Query Language is more structured
Corpus Query Language, or CQL, supports token-level conditions in compatible corpus software. Sketch Engine documents conditions such as [word="went"] for an exact word and [lemma="go"] for a lemma search. Attributes are properties supplied by the selected corpus.
CQL is not a universal standard accepted by all search engines. The corpus must carry the requested attributes, and grammatical tag names vary. The correct approach is to read the chosen tool’s documentation and begin with a simple working example before adding complexity.
11. How a CQL sequence works
In Sketch Engine’s documented syntax, successive bracketed conditions describe successive tokens. For example, [lemma="refill"] [lemma="the"] [lemma="teapot"] describes a sequence of three lemma matches. Spaces between bracketed conditions can be inserted for readability.
This illustrates how a search can express a structured condition rather than a single word. But an example in the documentation is not a measured finding about a corpus. We must run the query, identify the selected dataset and examine the returned sentences before making claims about frequency or usage.
12. Grammar tags narrow the search
Part-of-speech tags can distinguish nouns, verbs and other categories. If a student wants the verb record, including noun occurrences may interfere with the analysis. A tagged corpus can provide filters, although the exact tagset depends on the corpus.
Automatic labels are not always correct. A researcher should examine samples near ambiguous usages, quotations and unusual expressions. Good grammar searches help teachers ask a more precise question, but they do not replace grammatical interpretation of the returned sentence.
13. Metadata can make a subcorpus
Corpus metadata can describe the source, date, genre, author grouping or other details of texts. Filtering to a suitable subcorpus can make a classroom query more relevant. A student comparing formal instructions with conversational dialogue needs to know which examples belong to each set.
Filtering is not an automatic cure for biased data. If the sample consists mostly of one publisher or a repeated template, a genre label does not make it representative. Record what the subset contains, what it leaves out and why it was chosen for the particular learning question.
14. Context windows can mislead
A concordance line may display only a short span around the matched expression. That is convenient for scanning many examples but can conceal negation, reported speech or a condition elsewhere in the sentence. An isolated phrase can appear to support the opposite of what the author actually wrote.
For example, the sentence ‘The results did not support the claim’ must not be summarised as ‘The results support the claim.’ Expand to the full sentence or paragraph whenever the evidence would change with more context. Reading is part of querying, not an optional step after statistics.
15. Sorting a concordance uncovers candidates
Concordancers often let users sort examples by the immediate left or right neighbour of the searched word. This can reveal recurring prepositions, adjectives, verbs or phrases. The learner may discover a candidate expression worth checking against the intended meaning.
Sorting reorganises an existing result set; it does not prove that the most visible combination is correct in every context. Examine a range of examples, including less common ones, and check whether the same sense and grammatical relationship are present.
16. Filtering can become cherry-picking
After a search, a researcher may remove false matches or restrict the output to a genre. That can improve precision, provided the criteria are clear. The danger comes when inconvenient examples are deleted simply because they contradict the learner’s first guess.
Keep a simple query log: original search, filters, numbers of examples remaining and reason for each exclusion. If the analysis was exploratory, say so. A vocabulary rule becomes more reliable when it survives a fair comparison with counterexamples.
17. Counts depend on the query definition
An exact word search, lemma search and broad wildcard can yield different counts from the very same corpus. Tokenisation, punctuation handling and text duplication can also influence results. It is misleading to report a large total without identifying the matching operation.
A good school investigation records the corpus, query, settings and access date. If the examples are invented for teaching, do not fabricate hit counts and present them as empirical findings. Numbers become meaningful only when readers understand what was counted.
18. No match does not mean impossible English
A query returning zero examples could mean the phrase is rare, the corpus is narrow, the spelling differs or the query is incorrectly formed. It might also mean the corpus lacks a relevant sense or has unreliable annotation.
Check a reputable dictionary and consider broader, relevant sources before declaring a sentence ungrammatical. A careful conclusion is ‘We found no matches under these search conditions’, not ‘No English speaker ever uses this expression.’ The difference is a valuable lesson in reasoning.
19. False positives require sense checking
A false positive is an occurrence retrieved by the query that does not answer the intended question. Searching for financial charge may retrieve a battery-charging example. A surface-word query does not automatically know which of several meanings the reader wants.
Mark the sense of a sample of lines and remove irrelevant instances using a documented criterion. The pupil can identify clues in the surrounding sentence that distinguish the senses. This teaches context selection and gives the vocabulary investigation a defensible scope.
20. False negatives require alternative forms
A false negative is a relevant occurrence the search fails to retrieve. Looking only for analyse might miss analysed or a different spelling such as analyze. A narrow phrase query could overlook an inserted adverb that preserves the same function.
Widen the query methodically, but do not change the research question without noticing. Record each variation and its purpose. The goal is not to collect the highest possible number of examples; it is to collect the ones needed to answer the question accurately.
21. Precision and recall are different
Precision is the share of retrieved examples that actually match the learner’s purpose. Recall concerns how many relevant examples the search finds out of those available in the collection. A narrow query may give cleaner results yet miss alternatives; a broad query may include more useful possibilities alongside noise.
For a Primary composition correction, high precision and a few clearly understood sentences may be enough. For a research claim about the full range of an expression, recall becomes important. Choose the balance based on the stakes instead of treating every search as equally demanding.
22. Test simple conditions before complicated ones
A complex query can look scientific while containing a basic mistake. Begin with one word or phrase, inspect known examples and add constraints gradually. If the new condition is supposed to exclude nouns, check an example that is definitely a noun.
This is the same discipline used in Mathematics and computing: specify a rule, run a test and explain the outcome. When the search behaves unexpectedly, change one condition at a time. The student learns to debug a linguistic procedure rather than merely repeat it.
23. Case sensitivity is a design choice
A case-insensitive search may mix May the month with may the modal verb. A case-sensitive search may separate those forms but still encounter sentence-initial capitals and names. Neither switch perfectly identifies grammatical meaning.
When the purpose is to study permission or possibility, a grammatical or contextual classification may be necessary. Examine what the software retrieved rather than trusting that case alone identifies the intended sense.
24. The corpus itself can contain errors
Even large corpora contain spelling errors, unusual quotations, jokes, dialect forms, copied material and deliberate artistic language. Finding an expression in a dataset does not automatically make it suitable for an academic essay or a public safety notice.
A corpus answers where a form appears under a particular search rule. Dictionaries, source quality, communicative purpose and human interpretation still determine whether it belongs in a student’s sentence. Never use one search hit as a universal permission slip.
25. Primary English: word senses on paper
A Primary teacher can prepare ten original sentences containing light as a noun, adjective or verb. The learner circles only the instances referring to illumination, then explains the clues in each sentence. This is an invented teaching exercise, not corpus evidence.
The activity introduces the logic of a query without requiring software. Next, the child writes a new sentence for a different sense. The useful outcome is the ability to recognise and choose meanings, not memorisation of technical search terminology.
26. Secondary English: improve a composition
Imagine a student writes, ‘She performed a decision.’ The teacher first asks what actually happened. Was the decision made, announced, followed or implemented? Those actions are not the same. Authentic examples of decision can show possible grammatical partners.
The correction should preserve the story. If the character made a choice but has not acted, changing the verb to implemented would distort the event. Corpus queries help with form and convention, while the writer’s intention remains essential.
27. Mathematics: operational vocabulary
Words such as at least, at most, exactly and remaining can determine the operation a Mathematics problem requires. Students may use examples to become familiar with the wording, but the actual problem statement establishes the rule.
For instance, ‘at most five’ permits five, while ‘less than five’ excludes it. A language investigation should support the precise inequality rather than replace mathematical reasoning with a general impression of how the phrase is used elsewhere.
28. Science: keep evidence and inference apart
Science reports distinguish observed, measured, inferred and proved. A corpus query can reveal different grammatical contexts, but it cannot validate an experiment that was not performed.
The student must choose a verb consistent with the evidence. Seeing a colour change is an observation; claiming a specific cause may require further evidence. Vocabulary improves Science learning when it prevents the language from making stronger claims than the data support.
29. History: words can carry a viewpoint
A historical account may call an event a reform, restriction or concession. Corpus evidence can help explain connotations and source framing, yet the word itself does not reveal every motive or verify the historical interpretation.
Read the actual source, its author, date and audience; then corroborate with other evidence. The search supplies examples of expression, while source evaluation determines what conclusions are defensible.
30. Translation: preserve the communicative act
A direct word-for-word translation may change the politeness or force of a request. A target-language corpus can help investigate expressions used in comparable situations, provided its register and audience fit the task.
Identify what the speaker intends first: a request, warning, suggestion or command. Then examine candidate expressions in full contexts and choose one that preserves that function. Counting similar strings is not enough to establish equivalence.
31. Policy and workplace vocabulary
A workplace procedure may distinguish reviewed, acknowledged and approved. Archived correspondence can show examples, but the organisation’s current policy defines the authority behind each action.
Search tools may help identify inconsistent wording, yet they cannot manufacture an approval or a completed review. Accurate vocabulary requires checking what has actually happened, who has authority and which conditions apply.
32. Keep student data private
Building a class corpus from genuine student writing can reveal recurring teaching needs. It can also expose student names, school records, private incidents and personal details if copied into an external tool without adequate safeguards.
Use public licensed resources or clearly labelled invented exercises by default. For authorised research, follow the school’s requirements, limit access, remove identifiers where appropriate and describe the sample honestly. The benefits of analysis do not excuse careless handling of children’s writing.
33. AI can suggest searches but cannot claim they ran
A generative AI can help brainstorm possible queries or compose hypothetical example sentences. Unless it actually executed a request against a named corpus and made the results verifiable, its output is not evidence of language frequency.
A sensible workflow uses AI for scaffolding, the real corpus for observations and the student for interpretation. Generated exercises should be marked as invented. A convincing sentence or plausible count cannot stand in for provenance.
34. Exploration and formal testing differ
A learner may discover an unexpected phrase and refine the query repeatedly. That is legitimate exploration. The mistake would be presenting the final discovered pattern as though it had been predicted before any evidence was inspected.
Label exploratory observations, preserve the query history and, where the claim matters, test it on a fresh source. This habit transfers to data science, research and critical thinking: the origin of an idea affects how confidently it can be reported.
35. A useful query notebook
One page can record the question, corpus, search string, matching unit, filters, candidate examples, rejected examples and final choice of wording. A separate line should note important limits, including unavailable sources or ambiguous senses.
The notebook makes the method reviewable without requiring specialist software. A teacher can ask the learner to justify each decision, and another reader can understand what happened. An auditable small exercise often teaches more than a large unexamined output.
36. A fifteen-minute learning cycle
In an illustrative tutorial, spend three minutes identifying the exact word-choice problem, three defining the query, four inspecting relevant examples, three revising the student’s sentence and two asking for a fresh sentence without help.
The timings are teaching suggestions, not measured claims of learning effectiveness. A younger learner may spend the whole session on five invented examples. What matters is transfer: the student should be able to use the selected expression correctly in a new context.
37. When a result is unexpectedly strange
If a search retrieves an implausible grammatical relationship, check whether the words belong to the same clause, the annotation is correct or the example comes from a quotation or a different sense. A striking result can be the beginning of a good question.
Resist the temptation to teach every surprising phrase as a fashionable new collocation. Compare with a learner’s dictionary and other appropriate examples. An anomaly is evidence to investigate, not a command to imitate.
38. Do not confuse frequency with correctness
A word may occur frequently because it is part of a template or a specialist field. An unusual but grammatical expression may be rare in a small corpus. Correctness, clarity, naturalness and frequency are related but distinct questions.
The right expression for a task must communicate the writer’s actual meaning. Sometimes the simplest familiar verb is the strongest choice. The purpose of a corpus query is to broaden informed choices, not to force students into an arbitrary ranking of words.
39. The final test is a better sentence
After the investigation, close the search interface. Ask the student to write a sentence for a particular audience, identify its intended meaning and explain one reason the chosen expression fits. A counterexample or alternative may reveal the boundary of the rule.
The achievement is not remembering what the letters CQL stand for. It is knowing how to move from uncertainty to a clear question, traceable language examples, a careful interpretation and a better decision. That is vocabulary as an intellectual tool.
A reusable eight-step investigation
- Identify a real sentence that needs a more accurate vocabulary choice.
- State the meaning or grammatical pattern to investigate.
- Choose a suitable corpus and document its scope.
- Begin with a simple query and verify sample matches.
- Inspect larger sentence context, part of speech and sense.
- Record filters, rejected examples and relevant limitations.
- Rewrite the original sentence while preserving its factual meaning.
- Write a fresh sentence for a different audience and explain why the expression fits.
Ten questions and answers
1. Is a corpus query a dictionary?
No. It retrieves occurrences from a selected collection; dictionary entries organise meanings and usage guidance.
2. Does one example prove a phrase is normal?
No. One occurrence illustrates a possible use, not a general distribution.
3. What if my search finds no matches?
Check the dataset, spelling, query type and alternatives before interpreting the absence.
4. Can I use AI-generated sentences?
Yes, as explicitly constructed practice examples, not as measured corpus evidence.
5. Does CQL work everywhere?
No. Its syntax and fields depend on compatible tools and the annotations of the selected corpus.
6. Why search a lemma?
To retrieve relevant inflected forms when the corpus provides dependable lemma annotation.
7. Can sorting by neighbouring words prove a rule?
No. It helps locate patterns that still need contextual checking.
8. What is a false positive?
A retrieved instance that does not match the intended sense or condition.
9. Why keep a query log?
It makes the evidence, options and decisions understandable and reproducible.
10. What should a student take away?
The ability to test a word choice and write a new, accurate sentence.
Evidence sources and connected reading
- Sketch Engine: Concordance Guide — search types, filtering and reviewing results.
- Sketch Engine: CQL Basics — token conditions, lemmas and sequences.
- Sketch Engine: Practical CQL Examples — more complex corpus queries.
- Sketch Engine: CQL and Regular Expressions — search-language distinctions.
- AntConc: Concordance Guide — contextual reading and sorting in a free corpus tool.
- Vocabulary | What Is a Concordance Line? — understand the retrieved examples.
- Vocabulary | What Is a Corpus Comparison? — compare appropriately defined collections.
- Vocabulary Learning Hub — continue the wider learning pathway.
The search process matters because it disciplines our claims. Better questions and verifiable examples help people write with meaning, grammar and audience in mind. The goal is not the number of hits but the quality of the decision made from them.