Vocabulary | What Is a Collocation Window? How Word Distance and Direction Change What We Find

CHAPTER 1 OF 21

1. Begin with the target word rather than a score

A collocation analysis starts with a language question, not with a ranked list. The target word is usually called the node. A possible partner is a collocate, and the window says where that partner may appear. Without the node and window, a frequency figure has no fully specified meaning.

See the method in context

Suppose the node is the adjective strong. A learner wants to know which nouns frequently follow it. The relevant first question could be ‘Which noun occurs immediately to the right of strong?’ That is narrower than ‘Which words appear anywhere in sentences containing strong?’

The mistake hiding inside a plausible answer

The tempting shortcut reads a tool’s top-ten collocations without knowing what the tool counted. One interface may use nearby tokens, another grammatical relations, and a third a default symmetric span.

Make a better decision

Write the node, target part of speech if relevant, candidate category and window in one sentence. For example: ‘Find nouns within four tokens to the right of strong, remaining within the same sentence.’

Try this: What is missing from a list labelled simply ‘strong — top collocates’?

Explanation: The corpus, search target definition, span or grammatical relation, counting rules and score interpretation.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Next

CHAPTER 2 OF 21

2. Understand what left and right mean

A positional window is usually specified relative to the node occurrence. R1 means one token to the right; L1 means one token to the left. R4 reaches up to four rightward token positions, and L5 reaches up to five leftward positions. These values describe opportunities for a search, not physical distance.

See the method in context

In ‘Strong evidence supports the claim’, evidence is R1 from strong. In ‘The argument sounded strong’, argument is L2, because sounded occupies L1. The positions can be written on paper without specialist software.

The mistake hiding inside a plausible answer

The error calls any earlier word ‘the previous word’ or mistakes a direction indicator for a sentence-wide search. A collocation tool with R1 will not match a noun four tokens later.

Make a better decision

Number tokens around the node on both sides using a declared segmentation rule. Mark the exact offset for every candidate and then apply the window as a simple eligibility test.

Try this: In ‘The argument sounded strong’, where is argument relative to strong?

Explanation: L2 if tokens are The, argument, sounded, strong and punctuation is ignored for this illustrative analysis.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 3 OF 21

3. Separate window size from pair frequency

Window size sets eligibility; pair frequency records how many eligible node-candidate events were found. A wider right window usually provides more opportunities for a match but does not automatically create a distinct new observation from a candidate already within the narrow window.

See the method in context

A constructed set might show one occurrence of strong–evidence at R1, another at R2 and another at R4. An R1 query captures one event; an R4 query captures all three in those separate node contexts.

The mistake hiding inside a plausible answer

A misleading report says the corpus has changed because a wider window produces more results. The source text may be identical; only the rule describing what qualifies has changed.

Make a better decision

Keep raw pair counts next to the chosen window and count each matching event under the declared method. If a candidate appears twice near one node, specify whether both events or one per node are tallied.

Try this: Does a window of R4 always produce exactly four times the R1 frequency?

Explanation: No. Many eligible positions may contain other words or no token at a sentence boundary. The result depends on the actual text.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 4 OF 21

4. Choose immediate adjacency when grammar requires it

Some questions concern fixed-order local structures: an adjective directly modifying a following noun or a specific word pair with no intervening token. For those, an R1 or L1 analysis may be an informative first approximation when the relevant relation is genuinely adjacent.

See the method in context

‘Strong coffee’ is an adjacent adjective-noun pattern. A right-one query from strong captures coffee under a basic word-token policy. It will not capture ‘strong and carefully prepared coffee’, even if that expression conveys an adjective-noun relationship at a distance.

The mistake hiding inside a plausible answer

The mistake calls the R1 result a complete inventory of every phrase involving strong. It excludes legitimate structures separated by modifiers and may include adjacent words that are not grammatical collocates.

Make a better decision

Use adjacency for the specific question it can answer, then check whether relevant grammar permits intervening material. If it does, consider a larger window or syntax-aware query.

Try this: Can R1 measure all uses of strong modifying a noun?

Explanation: Not necessarily. Intervening words, punctuation and grammar can cause relevant relationships to fall outside strict adjacency.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 5 OF 21

5. See how inserted words change position

Adding a word between the node and candidate moves the candidate farther away, even if the writer’s broad meaning remains related. Positional distance measures the sentence’s visible token arrangement, not a stable semantic distance.

See the method in context

The sentence ‘Strong and carefully checked evidence matters’ places evidence at R4 from strong under the displayed segmentation: and is R1, carefully R2, checked R3 and evidence R4. A search limited to R3 misses it.

The mistake hiding inside a plausible answer

A shortcut assumes any word in the same short sentence is ‘nearby enough’. Window definitions are not intuitive judgments; they are precise boundaries that readers should be able to reproduce.

Make a better decision

Mark intervening tokens explicitly. If you want to study grammatical links across a coordinated or modified construction, consider whether simple proximity is adequate or a syntactic relation is needed.

Try this: At what smallest rightward window would this illustrative sentence include evidence?

Explanation: R4. It does not qualify for R1, R2 or R3 when the tokens are counted as specified.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 6 OF 21

6. Allow asymmetry when the question is directional

An L2R0 window examines only two positions before the node; an L0R4 window examines only four after it. Direction matters in languages where grammatical partners occupy characteristic positions, but it should follow the research question rather than a preference for larger counts.

See the method in context

An analyst studying adjectives before English nouns may begin with R1 from strong. An analyst studying words immediately before the verb explain might examine L1 or L2. Neither direction is universally correct.

The mistake hiding inside a plausible answer

A misleading comparison treats the L1 count and R4 count as equivalent measures of lexical preference. They have different eligible event spaces and may select different syntactic functions.

Make a better decision

Choose direction based on the construction being investigated. State exact left and right spans, not simply ‘five-word window’ when the choice could be ambiguous.

Try this: Does a ‘five-word window’ always mean five words total?

Explanation: No. Depending on the tool or description it might mean five on each side, five total, or a one-sided span. State L and R separately.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 7 OF 21

7. Treat punctuation as a processing decision

In many tokenisation schemes punctuation is handled as separate tokens; in others a word-only positional search ignores punctuation when counting word offsets. The choice can shift a candidate from R2 to R1 or produce different eligible events.

See the method in context

Consider ‘Strong, clear evidence matters’. If commas are ignored as word tokens, clear is R1 and evidence R2. If a particular tool counts the comma as a token, the offsets differ. Either analysis must be labelled.

The mistake hiding inside a plausible answer

A common error combines examples counted by one word-segmentation rule with a denominator or offset table produced by another. The window value looks identical while the effective distance differs.

Make a better decision

Check the corpus tool’s definition of token and whether the search operates on word tokens, all tokens or tagged positions. For classroom demonstrations, write the token list above the table.

Try this: Is evidence R2 in ‘Strong, clear evidence matters’ under the word-only rule used here?

Explanation: Yes. Clear is R1 and evidence R2 when the comma is not counted as a word token.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 8 OF 21

8. Decide whether a window may cross a sentence boundary

A word-based window can reach beyond a sentence ending unless the tool is instructed to stop at the boundary. For ordinary phrase analysis, cross-sentence proximity can create questionable pairings; for discourse studies, cross-boundary context may sometimes be useful.

See the method in context

In ‘The answer was strong. Evidence followed later’, a raw five-token window could place evidence close to strong even though it begins a separate sentence. Whether that should count depends on the declared analysis.

The mistake hiding inside a plausible answer

The shortcut assumes all tools silently respect sentence boundaries. Some search workflows require an explicit structural restriction. Without one, the output can mix local collocation with broader discourse proximity.

Make a better decision

Write the sentence or document boundary rule and test a constructed example that crosses it. If the corpus lacks reliable sentence segmentation, disclose that limitation rather than pretending the search was grammar-aware.

Try this: Should a phrase-level analysis silently count a collocate in the next sentence?

Explanation: Not without an explicit policy and rationale. Boundary handling determines what the co-occurrence result represents.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 9 OF 21

9. Keep document boundaries visible

Even a broad window should normally respect the unit of a language event when the aim is intra-text association. Joining separate documents into one stream can create artificial neighbours at file boundaries.

See the method in context

A school archive appends several notices without separators. The last word of one notice might appear next to the first word of another. Those words were not neighbours in the original communication.

The mistake hiding inside a plausible answer

The mistaken method computes collocations on a concatenated plain text without retaining document boundaries, turning archive order into linguistic proximity.

Make a better decision

Preserve source IDs and structure. Ensure that any window cannot cross a document boundary unless the task explicitly studies concatenated material, which would be unusual for natural collocation analysis.

Try this: Why does a file-boundary error matter even for a one-word window?

Explanation: It can create a false adjacent pair from the end of one document and the start of the next.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 10 OF 21

10. Separate proximity from grammatical relation

Two words can be adjacent without being directly linked by grammar, and grammatically related words can be separated by several intervening words. A positional window finds proximity; a word sketch or dependency relation attempts a more structured linguistic analysis.

See the method in context

In ‘a surprisingly strong argument’, strong and argument are adjacent, but a different phrase might place a modifier or coordinated material between two linked words. A broader window can retrieve candidates but may also admit unrelated ones.

The mistake hiding inside a plausible answer

The error assumes a larger span automatically equals deeper grammatical understanding. In fact it may increase thematic noise, selecting words that merely share a passage.

Make a better decision

Use positional searches when proximity itself is relevant and grammar-aware tools when the research question requires syntactic dependencies. In either case, inspect concordance examples to confirm the relationship.

Try this: Can L5R5 output be interpreted as an exact adjective–noun dependency list?

Explanation: No. It includes possible neighbours under a span; grammatical relation needs its own evidence or tagging.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 11 OF 21

11. Study repeated candidates near one node

A candidate may occur more than once within the same search window around a node. Different tools can count each occurrence as a pair, count the number of node contexts containing the candidate, or apply another event definition.

See the method in context

In the toy sentence ‘Strong evidence, stronger evidence and still strong evidence’, multiple occurrences and even multiple variants complicate what the word strong should contribute to one pair total. A simple one-line count is not self-explanatory.

The mistake hiding inside a plausible answer

A misleading table treats ‘found in four node contexts’ as identical to ‘four collocate-token events’ even when a context contains repeated candidates.

Make a better decision

Name the unit of the co-occurrence count. If a tool outputs an event count, document whether each node-candidate combination is counted once or repeatedly within a window.

Try this: How could one sentence contribute more than one pair event?

Explanation: It may contain multiple node occurrences or repeated candidate words within the eligible windows, depending on the counting policy.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 12 OF 21

12. See the effect of a wider window on topic words

A broad window can capture thematically connected words without capturing a reusable expression. It may be useful for studying semantic environments, but less suitable for teaching which words form a conventional phrase.

See the method in context

An examination report might mention students, assessment, marks and improvement within a few tokens of the word score. Those neighbours describe the topic even if they are not fixed collocations with score.

The mistake hiding inside a plausible answer

A teacher converts a wide-window list into sentence stems and produces unnatural phrases because the statistical neighbours were never evaluated as grammatical partners.

Make a better decision

Treat broad-span output as candidate context vocabulary, not a ready-made collocation dictionary. Read sentences, identify grammatical roles and consider narrower searches for instructional word partnerships.

Try this: Is a word occurring within five tokens necessarily a natural phrase partner?

Explanation: No. It can be a thematic neighbour with no direct grammatical relation to the node.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 13 OF 21

13. Work through a fully counted eight-sentence set

The constructed workbook contains eight sentences, each with one occurrence of strong. Evidence appears at R1, R4, R2 and L5 in four different sentences. Coffee appears at R1 and L3; argument appears at R1 and L2. The table records exact positions under a word-only, within-sentence policy.

See the method in context

An R1 search counts one evidence, one coffee and one argument event. An R4 search counts three evidence, one coffee and one argument. Adding L5 yields a symmetric-as-declared L5R4 view with four evidence, two coffee and two argument events.

The mistake hiding inside a plausible answer

The mistaken interpretation treats R1 and L5R4 frequencies as if they came from different corpora or as if an expanded search inherently proves one phrase is more meaningful.

Make a better decision

Reconstruct each position from the sentence tokens. Then apply the declared windows without changing the source text. This lets a learner see precisely why more hits appear.

Try this: How many strong–evidence events appear under R4 in this constructed set?

Explanation: Three: the R1, R2 and R4 cases. The L5 example is excluded until a leftward window is added.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 14 OF 21

14. Know that the statistical expected count also changes

An association statistic compares observed events with some expected or reference quantity under a model. Changing the size or direction of the co-occurrence window changes the eligible opportunities, so a statistic computed for R1 cannot be pasted unchanged onto an L5R4 result.

See the method in context

A collocation calculator may request corpus token count, node frequency, candidate frequency and separate left/right window sizes. The chosen span can affect how expected co-occurrence is modelled; the exact formula varies by tool and measure.

The mistake hiding inside a plausible answer

A serious mistake replaces the observed R1 count with the wider count but keeps an expected frequency computed for adjacency. The reported score then mixes incompatible measurement rules.

Make a better decision

Recalculate both observed and expected quantities under the same documented method. Use the calculator’s definition rather than assuming that every measure scales linearly with window length.

Try this: Can you reuse an R1 mutual-information score after switching to R4?

Explanation: No. The observed relation and the relevant expected model must be evaluated under the new window policy.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 15 OF 21

15. Interpret the size of the corpus and rare matches

A wide window often produces more candidate matches, but those matches may be concentrated in a few files or dependent on repeated text. A small corpus can produce unstable rankings, especially for unusual collocates.

See the method in context

An eight-sentence classroom exercise reveals the logic but cannot show which partners of strong are most typical in English globally. A larger corpus offers more observations, yet source coverage and duplicate texts still matter.

The mistake hiding inside a plausible answer

The false conclusion says one invented set proves a language-wide preference. A small constructed table teaches the meaning of a window, not the actual frequency of natural English expressions.

Make a better decision

Report raw node count, pair count, corpus domain and window. Where relevant, also show how many distinct documents contain each pair and inspect representative context examples.

Try this: Does a larger window solve a small-sample problem completely?

Explanation: No. It changes the event definition and may introduce more noise; additional representative material and careful validation may still be needed.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 16 OF 21

16. Explain why different tools disagree

Two systems may use different tokenisation, lemmatisation, sentence segmentation, direction conventions or default windows. They can produce different counts without either suffering a software failure.

See the method in context

One interface searches exact strong within three words both sides; another searches the adjective lemma STRONG within an R1 noun relation. These are not the same queries even if both display a column labelled collocate.

The mistake hiding inside a plausible answer

The error concludes a researcher fabricated results because two numbers differ, without comparing tool settings. Cross-tool validation requires methodological equivalence.

Make a better decision

Document the corpus version, target representation, L/R values, boundary policy, candidate category, counting unit and score. Reproduce one sentence manually to make the settings concrete.

Try this: When do two collocation totals become fairly comparable?

Explanation: When their source materials, search targets, event definitions and relevant calculations are compatible or their differences are explicitly accounted for.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 17 OF 21

17. Keep word-sense distinctions in view

The adjective strong has several uses: physical strength, concentration, evidential support and intense flavour. Even if a window reliably captures neighbours, one statistical pattern may mix senses unless the analysis inspects context.

See the method in context

Strong coffee and strong evidence are both ordinary English combinations, yet they describe different properties. A word-partner exercise should teach the meaning shift and why substituting powerful or intense may alter naturalness.

The mistake hiding inside a plausible answer

A weak method treats all partners as synonyms or concludes that the highest-scored partner expresses the only correct sense. Association does not settle meaning by itself.

Make a better decision

Group concordance lines by interpretable senses when the task requires it, and review exceptions. Teach the phrase and its communicative role rather than simply drilling a list of candidate nouns.

Try this: Can a high collocation score explain which meaning of strong applies?

Explanation: It can provide a clue but not a complete semantic account. Sentence context and grammatical function remain important.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 18 OF 21

18. Choose window settings for a writing lesson

A school writing lesson might begin with adjacent adjective-noun combinations because they are easy to observe, then expand the window to show modifiers, coordination and related grammatical patterns. The goal is accurate phrase choice, not an impressive statistic.

See the method in context

Students can inspect strong evidence, strong and carefully checked evidence, and evidence that is strong. They can explain how word order changes without losing the semantic link, and then practise writing a natural new sentence.

The mistake hiding inside a plausible answer

An unhelpful lesson asks learners to memorise L5R5 and a score threshold without ever constructing a sentence or discussing why a phrase is appropriate.

Make a better decision

Use a two-stage sequence: first identify the exact partner and position; next check grammatical and semantic relationship. End with retrieval and sentence-level correction.

Try this: What skill should remain after students forget the term L5R4?

Explanation: The ability to inspect context, recognise when words really form a useful partnership, and avoid overinterpreting proximity counts.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 19 OF 21

19. Separate exploratory and confirmatory choices

Researchers sometimes try several windows to discover where a pattern appears. Exploration is legitimate, but choosing only the setting that makes a desired contrast largest can exaggerate evidence.

See the method in context

Imagine checking L1R1, L3R3 and L5R5 and publishing only the last because it gives a high association score. Without a search log, readers may think the span was chosen before examining the result.

The mistake hiding inside a plausible answer

The mistake presents a tuned parameter as though it were an objective property of the word pair. The observed contrast partly depends on the selected measurement design.

Make a better decision

Declare an initial window based on linguistic rationale; if alternatives are explored, report them as sensitivity checks. Verify important patterns in comparable independent data when possible.

Try this: Why should an analysis disclose several attempted windows?

Explanation: It shows how sensitive the conclusion is to the analyst’s choice and helps prevent cherry-picking.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 20 OF 21

20. Build a paper-based quality check

Even without a corpus tool, a class can count token positions in a small text and inspect the effect of changing L and R. The practice is valuable because it exposes the same choices hidden inside a graphical interface.

See the method in context

Give learners eight printed constructed sentences and the node strong. Have them mark each candidate position and draw separate columns for R1, R4 and L5R4. Then compare counts with the supplied workbook.

The mistake hiding inside a plausible answer

The false shortcut marks any co-occurring sentence as a hit for every window. This ignores distance and direction, exactly the properties the activity is designed to teach.

Make a better decision

Require learners to justify every included pair by naming the position and sentence ID. Compare decisions and correct any disagreement at the tokenisation or boundary rule rather than guessing.

Try this: What is a useful answer format for an R4 hit?

Explanation: For example: ‘Sentence S02, evidence at R4 from strong, so this pair qualifies under R4 but not R3.’

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Next

CHAPTER 21 OF 21

21. Finish with an eight-part collocation-window audit

Before quoting a ranked collocation, verify corpus domain, target form or lemma, candidate definition, left/right span, tokenisation, sentence and document boundaries, event-count policy and score. These choices determine what the printed statistic can mean.

See the method in context

A cautious finding reads: ‘Under the constructed word-only within-sentence R4 search, evidence appears to the right of strong in three of eight target-node sentences.’ This is a checkable description of the exercise.

The mistake hiding inside a plausible answer

The overclaim says ‘strong evidence occurs three times more frequently in English than strong coffee’ from the toy R4 counts. The data are invented and the comparison omits genre, corpus coverage, independence and broader usage.

Make a better decision

Reproduce one row, compare narrow and wide windows, and record which claim survives. A score is only helpful when a reader can trace it back to a defined observation.

Try this: What transforms a collocation count into useful language evidence?

Explanation: A declared search method, inspectable examples, appropriate interpretation and explicit limits.

A reader can usually strengthen the result by asking three questions: Which source material was eligible? Which item was actually counted or classified? What is the smallest conclusion justified by that evidence? Write those answers explicitly before claiming that a pattern applies to every learner, writer, genre or period. When the evidence is too narrow, a more careful explanation is a better result than a dramatic statistic.

Contents · Previous · Worked table

Worked data: all decisions visible

This complete fictional eight-sentence example treats punctuation as non-word symbols, counts word tokens, uses exactly one node occurrence strong per sentence and never crosses a sentence boundary. R means positions to the right, L means positions to the left. The candidate is the noun evidence, coffee or argument. Each row contains one candidate and its position.

ID	Original constructed sentence	Candidate	Position from strong	R1 eligible?	R4 eligible?	L5R4 eligible?
S01	Strong evidence supports the claim.	evidence	R1	Yes	Yes	Yes
S02	Strong and carefully checked evidence matters.	evidence	R4	No	Yes	Yes
S03	Strong, clear evidence matters.	evidence	R2	No	Yes	Yes
S04	Evidence for this finding remains strong.	evidence	L5	No	No	Yes
S05	Strong coffee cooled quickly.	coffee	R1	Yes	Yes	Yes
S06	Coffee smells remarkably strong today.	coffee	L3	No	No	Yes
S07	The teacher made a strong argument.	argument	R1	Yes	Yes	Yes
S08	The argument sounded strong.	argument	L2	No	No	Yes

Under R1, the set has three eligible pair events: strong–evidence once (S01), strong–coffee once (S05), and strong–argument once (S07). The other sentences contain the candidate too far away or on the left.

Under R4, the rightward eligible count rises to five: evidence three times (S01, S02, S03), coffee once (S05) and argument once (S07). Both the text and number of strong nodes remain unchanged; the counting boundary has expanded.

Under L5R4, the counts are evidence four (S01–S04), coffee two (S05–S06), argument two (S07–S08): eight eligible pair events in total, three more than R4. Every row now qualifies. The difference between R4 and L5R4 comes solely from adding leftward positions in S04, S06 and S08.

The table supports counting positions, not claims about population frequency, significance or idiomaticity. An authentic tool would also require consistent tokenisation, boundary decisions and the intended corpus. Some tools differ on whether punctuation contributes to positional distance; here we explicitly use word-only offsets.

A quality-control exercise is to re-tokenise S02 as strong / and / carefully / checked / evidence / matters. Evidence lies at R4. Re-tokenise S04 as evidence / for / this / finding / remains / strong. Evidence lies at L5. This allows a reader to verify the most distant right and left examples manually.

Twelve transfer cases: the limits of an attractive number

Each scenario is an original classroom-style illustration. Read the case, identify what the analyst can actually observe, and compare the suggested repair with the evidence needed to carry it out.

1. Added modifier

Situation: ‘Strong evidence’ becomes ‘strong and carefully checked evidence’.

Hidden danger: An adjacent search misses a related word farther right.

Reasoned response: Use R4 if the question concerns a wider positional span, and verify grammar.

Next question: Which tokens occupy R1–R3?

2. Left-side evidence

Situation: A sentence ends ‘remains strong’ after mentioning evidence.

Hidden danger: R1 cannot see the earlier noun.

Reasoned response: Use a leftward window if the question warrants it.

Next question: What is the candidate’s exact left offset?

3. Sentence crossing

Situation: A target word ends one sentence; the candidate begins the next.

Hidden danger: A raw token window may count an unintended relation.

Reasoned response: Apply a declared sentence-boundary restriction.

Next question: Would the research question ever justify crossing it?

4. Document concatenation

Situation: Several archived files are joined without separators.

Hidden danger: A false pair appears at the file boundary.

Reasoned response: Preserve document IDs and prohibit cross-document co-occurrence.

Next question: How could an R1 false positive arise?

5. Comma handling

Situation: Two programs count punctuation differently.

Hidden danger: Offsets differ despite the same printed sentence.

Reasoned response: Use consistent word-token conventions and report them.

Next question: Is comma part of the distance here?

6. Wider thematic noise

Situation: L5R5 finds score near marks and pupils.

Hidden danger: Topical neighbours are mistaken for reusable phrases.

Reasoned response: Check grammar and context separately from proximity.

Next question: Would an adjective–noun relation find the same items?

7. Rare score

Situation: A candidate occurs once with high MI.

Hidden danger: One match may be unstable.

Reasoned response: Show raw frequency and inspect sources before teaching it.

Next question: What other evidence would strengthen the claim?

8. Duplicated notice

Situation: The same phrase is copied across many school pages.

Hidden danger: Pair counts may reflect one repeating template.

Reasoned response: Audit document spread and source reuse.

Next question: How many independent originals remain?

9. Exact versus lemma

Situation: Tool A searches strong; tool B groups stronger, strongest.

Hidden danger: Different node definitions change opportunities.

Reasoned response: Use compatible searches or disclose the difference.

Next question: Is the teaching question about one spelling or a lemma?

10. Grammar link

Situation: An adjective modifies a noun several tokens later.

Hidden danger: A strict R1 misses a syntactic relationship.

Reasoned response: Use appropriate grammar-aware analysis if the link matters.

Next question: Can proximity alone certify that relation?

11. Unequal window names

Situation: A paper says ‘five-word window’ without a direction.

Hidden danger: Readers cannot reproduce the exact span.

Reasoned response: Specify L and R values separately.

Next question: Does five mean total or per side?

12. Parameter tuning

Situation: An analyst tests many spans and publishes the highest score.

Hidden danger: Selection after results can inflate apparent certainty.

Reasoned response: Disclose exploration and validate on independent data.

Next question: Was the primary span predeclared?

Eight practice questions and explained solutions

Answer before reading the solution. Show raw evidence and units for a calculation; for a qualitative decision, state the scope and the reason. The purpose is transferable judgement rather than memorising a list of answers.

Exercise 1: Count word offsets

Question. In ‘Strong and carefully checked evidence matters’, where is evidence?

Explained answer. R4 under the word-only convention. Three tokens— and, carefully, checked — intervene.

Why the shortcut fails. A guess of R3 counts intervening words rather than the target’s position.

Exercise 2: Read left offsets

Question. In ‘Evidence for this finding remains strong’, where is evidence?

Explained answer. L5: remains is L1, finding L2, this L3, for L4 and evidence L5.

Why the shortcut fails. Sentence length alone is not the offset.

Exercise 3: Compare R1 and R4

Question. Use the workbook: how many strong–evidence pairs qualify under R1 and R4?

Explained answer. R1: one (S01). R4: three (S01–S03). The fourth evidence example is on the left.

Why the shortcut fails. Treating all same-sentence matches as rightward ignores direction.

Exercise 4: Count a two-sided window

Question. How many evidence, coffee and argument events appear under L5R4?

Explained answer. Four evidence, two coffee and two argument, eight in total.

Why the shortcut fails. R1-only counting would miss five of these eligible events.

Exercise 5: Protect sentence boundaries

Question. ‘Strong. Evidence appeared later.’ Should a phrase-level within-sentence window count the pair?

Explained answer. No, because the evidence lies outside the same sentence under that explicitly chosen policy.

Why the shortcut fails. A raw token window may cross punctuation unless constrained.

Exercise 6: Interpret ranking shift

Question. Why might a pair’s MI change after increasing the window?

Explained answer. The observed co-occurrence and modelled opportunities can change. Recalculate under the declared statistic and span.

Why the shortcut fails. Keeping the old expected count while changing O mixes methods.

Exercise 7: Distinguish theme from phrase

Question. A wide window finds hospital near doctor. Is that necessarily a fixed two-word expression?

Explained answer. No. It can be a semantic or topical association without forming a particular grammatical phrase.

Why the shortcut fails. Proximity alone is not proof of syntactic partnership.

Exercise 8: Write a methodological caption

Question. Write one sentence describing the R4 result for evidence in the invented workbook.

Explained answer. ‘In eight constructed sentences with strong as the node, evidence falls within the four rightward word positions in three sentences under a word-only, within-sentence counting rule.’

Why the shortcut fails. ‘Evidence is three times more common in English’ would be an unsupported generalisation.

Frequently asked questions: Collocation window

What is a collocation window?

The positional range or structural span in which a corpus search counts candidate partners around a target word.

What do L1 and R1 mean?

One token to the left and one token to the right of a node, under a declared tokenisation policy.

Is L5R5 the same as five words total?

No. It commonly describes five positions on each side; always check the tool’s exact convention.

Does a wider window always produce a better collocation list?

No. It can retrieve useful distant patterns but also increase thematic noise.

Does changing the window change the underlying corpus?

Not necessarily. It changes the qualifying co-occurrence events and often the scoring opportunities.

What happens at sentence boundaries?

A tool may require an explicit structural restriction. State whether the window stops or crosses.

Does punctuation count as a token?

It depends on segmentation and the software’s search semantics; document the rule.

What is the difference between co-occurrence and grammatical relation?

Co-occurrence describes proximity under a span; grammatical relation describes a syntactic link and may need annotated analysis.

Can left and right spans be unequal?

Yes. Directional questions often justify an asymmetric window such as L1R4.

Why might two corpus tools show different collocates?

They may differ in tokenisation, lemmas, spans, boundaries, candidate restrictions or ranking formulas.

Must I recalculate association scores after changing the window?

Yes. The observed events and expected model have to match the new counting rule.

Should a collocation list always be used directly for writing instruction?

No. Inspect authentic sentences, grammar, meaning and register before teaching expressions.

Can students learn the concept without software?

Yes. Number word positions in a small set of sentences and compare which events qualify under different windows.

How do I report a defensible result?

State the corpus, target definition, L and R span, segmentation, boundary rule, event count and relevant limits.

Compact vocabulary glossary

• Node: Target item whose neighbours are examined.
• Collocate: A candidate word meeting the chosen co-occurrence rule.
• Window: The positional span around the node.
• R1: One word position to the right.
• L1: One word position to the left.
• Co-occurrence: An observed node-candidate event satisfying a specified condition.
• Tokenisation: Division of text into countable units.
• Concordance: Word examples displayed with context.
• POS tag: A part-of-speech annotation used to restrict searches.
• Dependency relation: A grammatical connection between items.
• Association score: A statistic describing how a pair behaves relative to a specified model.
• Document frequency: How many distinct documents contain a given observed pattern.

A classroom routine for learning the method

Begin with one surprising example and have students make a prediction before revealing its count or classification. Ask them to state the rule that would settle the question. A learner who can make and defend a prediction is doing more than remembering a definition: the learner is reasoning from language evidence.

Next, introduce a counterexample. Change a short phrase’s word order, add a negation, move one noun beyond the selected position, or alter the sampling mixture. Hold everything else constant. Let students identify which conclusion becomes invalid, and why. This deliberate small change makes the scope of a rule visible.

Then ask students to recreate a compact evidence table without copying the supplied answers. They should name each column, check one calculation or label independently, and write a caption saying precisely what the table represents. Invite a peer to test whether the reported number follows from the chosen method.

Finally, apply the idea to a new reading or writing task. A good transfer exercise ends with two sentences: one specific observation supported by the materials, and one limitation that would matter when applying the result elsewhere. Revisit the exercise after a delay to test retrieval, and correct a mistake without hiding the original reasoning.

Methodological reading and verification

• Lancaster University: BNCweb collocations and window-size settings
• Lancaster corpus statistics tools: collocation calculator with left/right windows
• WebCorp LSE: collocates by left and right position and document frequency
• Sketch Engine: word sketch and grammatical collocations

The sources support the methods described. Illustrative counts and example messages are original to this article and should not be cited as empirical study results.

Continue through connected vocabulary guides

• Vocabulary | What Is a Collocation Association Score?
• Vocabulary | What Is a Word Sketch?
• Vocabulary | What Is a Concordance Line?
• Vocabulary | What Is a Subcorpus?
• Vocabulary | What Is a Corpus Sampling Frame?
• Vocabulary | What Is Normalised Word Frequency?
• Vocabulary Learning Hub
• Vocabulary articles index

Return to contents · Vocabulary Learning Hub