Collostructional Analysis¶
A corpus method for measuring lexical preference for a grammatical-construction slot against a frequency baseline.
Core Idea¶
Collostructional analysis is a family of corpus-linguistic procedures for asking which lexemes preferentially fill a slot in a specified grammatical construction. It combines a linguistic decision—what counts as an instance of the construction and its slot—with a distributional comparison: how often a word appears there relative to its occurrence elsewhere, or relative to its occurrence in a competing construction. The result is a ranked profile of attracted, repelled, or distinctive collexemes; that profile can inform, but does not mechanically determine, an account of constructional meaning.[1][2]
In the original single-construction procedure, each candidate lexeme is compared using four corpus counts: lexeme in target construction, other lexemes in target construction, lexeme outside it, and other lexemes outside it. Stefanowitsch and Gries favored Fisher's exact test for sparse counts; the named method is not restricted to that one statistic. In the distinctive variant, the two rows instead represent two competing constructions. The numerical association and the direction of preference must be read together: a small p-value alone does not say which construction attracts the word.[1][2]
Structural Signature¶
Sig role-phrases: delimited constructional slot → candidate collexeme and corpus → four-cell comparison → association plus direction → ranked lexical profile → guarded semantic interpretation.
- Construction and slot: a grammatical pattern, such as the noun slot of [N waiting to happen] or the verb slot of an into-causative, is identified before counting. Removing the constructional unit turns the method into generic collocation or word-frequency analysis.[1]
- Collexeme and corpus: the candidate lemma is counted in a declared body of texts. Its overall frequency matters: a common word may appear often in a slot without preferentially belonging to it.[1]
- Contingency baseline: four cells cross the candidate versus other lexemes with the target versus reference context. For single-construction analysis the reference is outside the construction; for distinctive analysis it is the competing construction. These are related but not interchangeable baselines.[1][2]
- Association and direction: a suitable test quantifies departure from a frequency-based expectation; observed-versus-expected occupancy identifies attraction or repulsion, or which alternative is preferred. Statistical significance by itself has no direction and is not identical to a substantive effect size.[2]
- Ranked profile and interpretation: repeating the comparison across collexemes yields a distributional profile. The original papers use it as evidence about constructional meaning, but one lexeme's score is not a complete semantic definition.[1][2]
- Operationalization boundary: corpus coverage and theoretically justified decisions about which tokens count as construction instances condition the result. Changing those decisions can change the table without refuting the method.[1][2]
What It Is Not¶
Collostructional analysis is not a list of frequent words near another word. Ordinary linear collocation does not require identifying a grammatical construction and one of its slots; this method does. Nor is raw frequency within a construction sufficient. The observed count must be interpreted against a comparison that accounts for how common the word and construction are overall.[1]
It is not the Fisher exact test itself. Fisher is a statistical instrument used in the original implementations; the research design also requires selecting the construction, extracting relevant tokens, constructing the right frequency table, determining direction, and interpreting a profile. A tiny p-value is not “the amount of meaning” or an effect-size unit. In the 2004 distinctive analysis, the authors explicitly use observed versus expected counts to determine which construction is favored after finding a very small p-value.[2]
It is not proof that every word in a ranked list is semantically central, or that two constructions are wholly synonymous or wholly distinct. Corpus representation, token coding, sparse items, and alternative explanations remain matters for linguistic judgment.[1][2]
Scope of Application¶
The single-construction setting asks, for a fixed pattern, which lexemes its slot attracts or repels. Stefanowitsch and Gries analyze the noun slot of [N waiting to happen] and the verb slot of the into-causative. In their corpus, accident and disaster rank strongly for the former; verbs such as trick, fool, coerce, and force rank for the latter. These are findings under the authors' corpus and coding, not universal word-to-construction laws.[1]
The distinctive-collexeme setting compares constructions that can fill similar communicative roles. Gries and Stefanowitsch compare the English ditransitive (sent Mary the book) with the to-dative (sent the book to Mary) and ask which verbs prefer which form. Here the reference class is the alternate construction, not simply all other corpus contexts; that choice changes what the statistic answers.[2]
Related co-varying analyses can ask about lexical preferences between two slots of one construction, but the present source-grounded examples and claims concern the single-construction and distinctive pair variants. A claim about a third variant should not borrow these two contingency baselines without specifying its own.
Clarity¶
The method separates three propositions often conflated in casual corpus discussion: a word occurs in a slot, it occurs there more than a frequency baseline predicts, and the set of preferential words helps characterize the construction. Each requires another step. A rare word's 14 occurrences may be informative if its corpus frequency predicts far fewer; a very common word may occur hundreds of times without being relatively preferred.[1]
It also clarifies the direction of a distinctive result. In the 2004 dative example, give appears in both constructions. Its Fisher p-value reports a strong departure from the null allocation, but direction comes from the comparison: 461 ditransitive occurrences versus about 213 expected, and 146 to-dative occurrences versus about 394 expected. The result is a ditransitive preference, not mere evidence that the two forms differ in some unspecified way.[2]
Manages Complexity¶
A corpus may contain thousands of candidate lexemes and many lexicalized examples. The procedure reduces them to repeated construction-conditioned four-cell comparisons and an ordered lexical profile. This makes a hypothesized semantic restriction inspectable: instead of appealing only to a handful of memorable phrases, an analyst can show which words are more or less strongly associated under the same counting rules.[1]
The reduction does not erase methodological obligations. The profile is meaningful only with a defensible token universe, construction boundary, lemma policy, and comparison set. A ranking from a broad mixed corpus and one from a narrow genre may disagree because their denominators and uses differ. The method manages combinatorial lexical detail, not all uncertainty about language use.[1][2]
Abstract Reasoning¶
Choose a construction C and candidate lexeme L. Count L-in-C, non-L-in-C, L-outside-C, and non-L-outside-C. Compare the observed L-in-C cell with the value implied by the row and column totals under an independence baseline, then use an appropriate association statistic. If the observed cell is greater than expected, the direction is attraction; if lower, repulsion. Repeat for many L and examine the ranked semantic grouping rather than assigning meaning from one cell.[1]
For the distinctive variant, replace C versus outside-C with construction A versus construction B. The logic remains a lexeme-by-context comparison, but the question changes from “Does L prefer this pattern over its corpus background?” to “Which of these two patterns does L prefer?” In the give example, a highly significant table plus observed/expected counts supports a ditransitive preference. It does not say that give never appears in the to-dative, since 146 such tokens are in the authors' table.[2]
Knowledge Transfer¶
The method literally transfers within corpus linguistics from partly fixed constructions to abstract argument-structure or aspectual patterns when the analyst can delimit a construction and lexical slot. The 2003 paper applies one logic of comparison across several such linguistic levels, while acknowledging that the token-identification task changes.[1]
The distinctive variant transfers to another pair of grammatical alternatives by replacing the pair and rebuilding its counts; the dative case is a template for design, not a result copied to all alternations.[2] A generic customer-product association table outside language may instantiate a broader association-analysis pattern, but it is not collostructional analysis unless a linguistic construction and its lexical occupancy remain the object.
Examples¶
Accident in [N waiting to happen]. Mapped back: constructional slot = N in the specified pattern; collexeme and corpus = accident in the authors' British National Corpus extraction; four-cell baseline = 14 accident occurrences and 21 other nouns inside the construction versus 8,606 accident and 10,197,659 other cases outside; association and direction = the authors report Fisher p about \(2.12\times10^{-34}\), with attraction because the in-slot count exceeds the baseline expectation; ranked profile = accident and disaster head the list; interpretive limit = this supports an undesirable-event reading but does not itself settle every use of the pattern.[1]
Give in two dative constructions. Mapped back: constructional slots = verbs in ditransitive and to-dative tokens; collexeme and corpus = give in ICE-GB; four-cell baseline = ditransitive 461 give and 574 other verbs, to-dative 146 give and 1,773 other verbs; association and direction = Fisher p \(1.84\times10^{-120}\) plus 461 observed versus 213 expected in the ditransitive; ranked profile = give is highly distinctive for that form, while other verbs must also be examined to explain the two constructions' semantics; limit = the result is specific to the defined pair and corpus.[2]
Raw-frequency near miss. A list saying “force occurs 101 times in the into-causative” identifies a verb and slot but leaves out force elsewhere and the alternative verbs. It cannot by itself establish preferential attraction. The original analysis supplies the comparison and ranks force alongside trick, fool, and others.[1]
Structural Tensions¶
Construction specificity versus sample size. Narrow inclusion criteria keep the counted tokens linguistically comparable but may leave sparse cells; broad criteria provide more data while risking a mixture of distinct constructions. Both aims cannot be maximized by changing the boundary silently. Diagnostic: Would a proposed coding change alter which corpus tokens count as the same construction?[1]
Numerical significance versus semantic explanation. A test can precisely rank departures from a selected null model; meaning requires inspecting multiple collexemes and their contexts. Leaning only on p-values risks semantic overclaim, while impressionistic reading without a baseline loses the method's quantitative check. Diagnostic: Is the proposed meaning supported by a profile of lexical preferences and actual uses, or just by one small p-value?[1][2]
Single-pattern background versus pairwise contrast. A broad outside-C baseline answers whether a word is special to one construction against its corpus environment. A pairwise A/B baseline answers which of two specified constructions it favors. The former is broader but may dilute a narrow alternation; the latter is sharper for the pair but cannot describe preferences outside that pair. Diagnostic: Which reference class matches the linguistic question?[2]
Structural–Framed Character¶
Evaluative weight. Attraction and repulsion are technical direction labels, not praise or disapproval of a word. Human-practice dependence. Researchers choose a corpus, define the construction, and interpret the profile; once those are fixed, the counts and test have reproducible consequences. Institutional origin. The method emerged from construction-grammar and corpus-linguistic research, but its evidential result is not settled by an institution's authority.[1]
Vocabulary travel. Terms such as “attraction,” “association,” and “preference” travel easily, yet the named collostructional procedure depends on lexemes occupying grammatical slots. Import versus recognition. Applying it to a new language or construction imports an extraction and testing practice; noticing frequent co-occurrence without that procedure is not an instance.
Its character: a substantially structural method within linguistics, but domain-specific as a named abstraction. Its cross-setting range covers linguistic construction types, not arbitrary associated entities.
Structural Core vs. Domain Accent¶
Portable skeleton. Compare observed and baseline frequencies in a typed contingency table, determine direction, then cautiously interpret a ranked profile. That broad association-and-inference skeleton can occur elsewhere, but the Encyclopedia's live Statistical Inference and p-Value entries cover aspects of that wider formal layer rather than making this linguistic method a prime.[1][2]
Domain accent. A grammatical construction, slot, lexeme, corpus tokenization, and constructional-semantic interpretation are constitutive. Replacing them with arbitrary objects preserves only a higher-order statistical resemblance. The method's evidential force also depends on grammatical decisions about what counts as an instance.[1]
Why not prime. The waiting-to-happen and dative examples are distinct linguistic settings, not independent-domain realizations of the named procedure. The substrate-neutral comparison belongs to broader statistical abstractions. No broad prime identity is asserted from metaphorical “attraction.”
Instantiates / Related Primes¶
Statistical Inference and Statistical Significance (p-Value) are related live primes: collostructional analysis often reasons from corpus counts with a test statistic, but its corpus can be treated as the observed collection rather than always as a random sample from a specified population. The live Statistical Test is another related instrument, not automatically the whole method's direct genus.
Neighborhood in Abstraction Space¶
Collostructional Analysis sits in a moderately populated region (57th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.
Family — Language, Mind & Meaning-Making (57 abstractions)
Nearest neighbors
- Tagmeme — 0.86
- Factored Language Model — 0.85
- Kneser–Ney Smoothing — 0.85
- Poetic Metre — 0.85
- Topic Marker — 0.85
Computed from structural-signature embeddings · 2026-10-08
Not to Be Confused With¶
Linear collocation analysis tests nearby words without making a grammatical construction's slot the target. Raw concordance or frequency lists display observed tokens but omit the frequency baseline needed to call them preferential. Fisher's exact test evaluates a contingency table; it does not choose the construction, collect the corpus, or supply a semantic explanation. Construction Grammar is a theoretical framework compatible with the method, not the method itself. Distinctive-collexeme analysis is a specified pairwise variant of the family, not a completely unrelated procedure.[1][2]
References¶
[1] Anatol Stefanowitsch and Stefan Th. Gries, “Collostructions: Investigating the interaction of words and constructions”, International Journal of Corpus Linguistics 8, no. 2 (2003), 209–243, especially §2.2 Tables 3–4 and §3.2.1 Table 8. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u ↩v ↩w
[2] Stefan Th. Gries and Anatol Stefanowitsch, “Extending collostructional analysis: A corpus-based perspective on ‘alternations’”, International Journal of Corpus Linguistics 9, no. 1 (2004), 97–129, especially §2 and Table 1. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r