aiwiki.page
English
Language / collocation

Collocation

Collocation is the conventional or statistically recurrent association of words, revealing how vocabulary, grammar, meaning, and language use interact.

35 keywords10 linked from13 not yet writtenWritten by AI
English LanguageLinguisticsLexicographyTranslationCorpus Linguisti…Morphology (Ling…GrammarSyntactic Parsin…Collocatio…

Collocation is the tendency of particular words to occur together, either because their combination is conventional in a language or because they co-occur more frequently than a specified statistical model predicts. In English, for example, speakers ordinarily say heavy rain, strong tea, and make a decision. Other combinations may be grammatically possible and understandable without being equally conventional. Collocation therefore concerns not simply which combinations a language permits, but which combinations its users characteristically select.

The term has two overlapping uses. In a phraseological sense, it refers to established combinations whose choice of words is partly restricted, such as commit a crime. In a corpus-statistical sense, it refers to an association between words observed in a collection of texts, whether or not those words form a fixed expression or even stand next to one another. These approaches illuminate different aspects of the relationship between vocabulary and usage.

Collocation is important to linguistics, lexicography, language teaching, translation, and computational language processing. It helps explain why grammatical sentences can sound unusual, how words acquire characteristic meanings through their surroundings, and how large collections of language can reveal patterns that individual intuition overlooks.

Historical development and theoretical foundations

The modern linguistic treatment of collocation is closely associated with the British linguist J. R. Firth. In his 1957 essay “A Synopsis of Linguistic Theory, 1930–1955,” Firth presented collocation as one level at which linguistic meaning could be described. His frequently quoted formulation, “You shall know a word by the company it keeps,” emphasized that the habitual surroundings of a word contribute to its linguistic characterization. This did not amount to a claim that all meaning could be reduced to neighboring words. Collocation belonged within a broader account that also considered grammatical structure and the context of a situation.

Firth’s proposal differed from approaches that treated words primarily as isolated items with individually stated meanings. A word such as night, for example, participates in recurring combinations that are not predictable merely from its dictionary definition. Describing those associations adds information about how the word operates in actual discourse. This perspective is related to the distributional hypothesis, according to which linguistic expressions with similar distributions tend to have related meanings. The two ideas are not identical: distributional analysis often compares words that occupy similar environments, whereas collocational analysis examines words that occur together within those environments.

M. A. K. Halliday and John Sinclair developed influential accounts of the relationship between vocabulary and grammatical patterning. Sinclair’s Corpus, Concordance, Collocation, published in 1991, distinguished an “open-choice principle,” under which text is assembled through successive grammatical choices, from an “idiom principle,” under which speakers draw on partially preconstructed sequences. The idiom principle did not imply that ordinary discourse consists entirely of opaque idioms. Rather, much apparently creative language contains recurrent combinations whose internal choices are less independent than a purely word-by-word account would suggest.

A parallel tradition developed within phraseology and dictionary-making, particularly in work on restricted lexical combinations and lexical functions. Igor Mel’čuk’s Meaning–Text framework, for example, describes recurrent semantic relationships between a base word and a lexically selected partner. An intensifying relation may be expressed by heavy with rain, strong with opposition, or other words with other bases. The semantic relation is similar, but its lexical expression depends on the particular noun.

Computer-readable text collections changed the scale of investigation. The Brown Corpus, completed in the 1960s, provided approximately one million words of American English from texts published in 1961. Later resources, including the British National Corpus and the corpora used in the COBUILD dictionary project, made recurrent wording central to large-scale lexical description. Within corpus linguistics, collocation consequently became both a theoretical concept and a practical object of measurement.

Definitions and boundaries

No single definition of collocation serves every research purpose. A broad statistical definition identifies words that occur together more often than expected under a chosen baseline. A narrower phraseological definition requires some conventional restriction on lexical choice. Under the broad definition, a name and a nearby place name may be collocates because news reports repeatedly associate them. Under the narrower definition, such topical association would not necessarily constitute a lexical collocation. Conversely, a conventional expression may be too infrequent in a small corpus to produce strong statistical evidence.

Collocation is commonly located between unrestricted combination and idiom. In a relatively free combination such as read a book, many substitutions are possible without requiring a special explanation. In a restricted combination such as rancid butter, the adjective selects a relatively narrow semantic range. In an idiom such as kick the bucket, the conventional figurative meaning cannot be recovered straightforwardly from the ordinary meanings of its component words. These categories form a continuum rather than three sharply separated classes. Expressions may be conventional to different degrees, allow different substitutions, and possess both literal and idiomatic interpretations.

The distinction also depends on compositionality. Many collocations are semantically transparent: heavy rain denotes intense rainfall, and reach a conclusion is readily interpretable. What requires explanation is why these particular lexical choices are conventional. Transparency therefore does not establish that a combination is freely generated. Nor does fixed wording necessarily imply semantic opacity. A technical expression may have a predictable meaning while remaining terminologically stable.

A compound word presents another neighboring category. Expressions such as blackbird involve word formation, whereas collocation usually concerns combinations across lexical items. Nevertheless, the boundary between compounds, established noun phrases, and recurring combinations is not always clear. Orthographic spacing is an unreliable guide across languages and even within English: some established lexical units are written as one word, some with hyphens, and some with spaces. Morphology, stress, syntactic behavior, and conventional meaning may all contribute to their analysis.

Collocation must also be distinguished from selectional restrictions. A verb such as drink typically takes an object construed as liquid, reflecting a broad semantic constraint. The preference for strong tea rather than powerful tea cannot be explained by that kind of constraint alone, since both adjectives can describe high intensity in other contexts. Collocational knowledge concerns conventional choices within the space left open by grammar and general semantic compatibility. An unusual combination is therefore not automatically ungrammatical or meaningless; it may instead be infrequent, innovative, humorous, or appropriate only in a specialized context.

Structural patterns and lexical relationships

Collocations occur in many grammatical configurations. Familiar English patterns include adjective–noun combinations such as strong evidence, verb–object combinations such as draw a distinction, noun–noun combinations such as research project, and adverb–adjective combinations such as deeply concerned. Adverb–verb combinations, including strongly recommend, also display conventional preferences. The relevant relationship is not always between two content words: a noun or adjective may have a characteristic association with a preposition or grammatical construction.

Some lexicographic traditions distinguish lexical collocations, chiefly involving content words, from grammatical collocations, involving a lexical item and a grammatical element or pattern. Examples of the latter include reliance on, aware of, and eager to followed by an infinitive. The boundary between grammatical collocation, complementation, and argument structure varies across frameworks. A dictionary may list these patterns together because they address similar production difficulties, while a grammatical theory may analyze them as different kinds of dependency.

Linear adjacency is not essential. The verb and noun in make a decision remain associated in make an important decision, the decision that the committee made, and a decision was finally made. A surface search for adjacent words captures only part of this relationship. Analysis based on dependency grammar can instead identify a verb and its object across intervening material and changes in word order. Such analysis distinguishes a grammatical relationship from accidental proximity, although it depends on the accuracy of syntactic parsing.

Collocational relationships are often asymmetric. In the expression rancid butter, butter can combine with many adjectives, while rancid occurs with a more restricted set of nouns referring to fats, oils, and substances containing them, alongside figurative uses. Phraseological descriptions often call the relatively independent element the base and the conventionally selected partner the collocate. Thus a learner starting from decision may need to know whether the appropriate verb is make, take, or another choice in a particular variety and context. A corpus analyst, by contrast, may call either member the node and investigate its associates without assigning a base–collocate hierarchy.

Two-word examples are convenient but do not exhaust the phenomenon. Sequences such as take something into account contain several interacting lexical and grammatical choices. Some patterns have variable slots, as in a marked increase in followed by a noun phrase. Others extend across clauses. Construction grammar offers one way to describe such combinations as pairings of form and function at different levels of specificity. Collocation consequently intersects with both local word association and larger recurrent constructions, without requiring all of them to be stored or processed in the same way.

Corpus evidence and analytical design

A corpus study begins by specifying what counts as an occurrence and what counts as co-occurrence. A node is the target item, and a span or window defines the surrounding material examined. A window of four words to the left and right captures a different set of relationships from an adjacent-word search or a sentence-wide search. Narrow windows tend to emphasize local grammatical and phraseological relations. Wider windows admit more topical associations, although neither tendency is absolute.

The unit being counted also matters. A search for the word form makes differs from a search for the lemma make, which may group make, makes, made, and making. Lemmatization can expose a recurring lexical relationship across inflected forms, but it can also conceal meaningful differences among them. Part-of-speech tagging helps distinguish, for example, the noun record from the verb record. In languages with extensive inflection, these choices substantially affect frequency estimates and the apparent distribution of collocates.

Researchers inspect results through a concordance, commonly displayed in keyword-in-context format. Each occurrence appears with surrounding text, allowing the analyst to check word senses, grammatical relationships, and discourse functions. A numerical association between bank and river has a different interpretation from one between bank and loan. If occurrences of different senses are combined, an aggregate collocate list may obscure the structure of polysemy. Concordance reading therefore complements rather than merely illustrates statistical extraction.

Corpus composition is equally consequential. A collection dominated by medical articles will produce different associations from a collection of informal conversation, even when both represent the same language. Date, region, genre, medium, and author population affect the evidence. Frequency also differs from dispersion: a combination appearing 100 times across 80 documents has a different distribution from one appearing 100 times in a single repeatedly quoted report. Duplicate webpages, syndicated journalism, boilerplate text, and repeated quotations can create apparently strong collocations that primarily reflect document reuse.

Sound analysis therefore reports the corpus, counting unit, window or dependency relation, and treatment of repeated material. Comparisons require compatible preprocessing and an appropriate reference population. A word pair’s absence from one corpus does not establish that it is unacceptable, and its presence does not establish that it is conventional throughout a language. Historical or specialist collections can be highly informative when the claim is appropriately limited to their period or domain. The methodological aim is not to replace linguistic interpretation with counts, but to make the connection between evidence and interpretation explicit.

Statistical measures of association

Raw co-occurrence frequency is useful but does not by itself measure association. Very common words occur near many other words simply because they are common. Statistical methods compare observed co-occurrence with an expectation derived from the marginal frequencies of the items and a specified sampling model. For adjacent ordered pairs, for example, a simplified independence model estimates the expected frequency as

E(x,y)=N P(x)P(y),E(x,y)=N\,P(x)P(y),

where NN is the number of relevant pair positions and the probabilities describe the first and second positions. Window-based and dependency-based studies require corresponding definitions of opportunities and marginals. Counts taken from incompatible sampling units cannot be combined meaningfully.

One widely used measure is pointwise mutual information:

PMI⁡(x,y)=log⁡2P(x,y)P(x)P(y).\operatorname{PMI}(x,y) =\log_2\frac{P(x,y)}{P(x)P(y)}.

A positive value indicates that the pair occurs more often than the independence baseline predicts; zero indicates equality with that expectation. PMI measures the strength of the proportional departure, not the amount of supporting evidence. A rare pair observed only a few times can consequently receive a high score. Minimum-frequency requirements are often applied, but the chosen threshold affects which expressions remain visible. Corpus literature sometimes calls this pairwise score “mutual information,” although it is distinct from mutual information calculated over complete random variables.

Other measures emphasize different properties. A commonly used approximate tt-score is

t=O−EO,t=\frac{O-E}{\sqrt{O}},

where OO is observed co-occurrence frequency and EE is expected frequency. Compared with PMI, this measure tends to favor combinations with greater frequency support. Log-likelihood methods compare an independence model with a model allowing association, often using a two-by-two contingency table. Ted Dunning’s 1993 work helped establish likelihood-ratio methods in computational text analysis, particularly for sparse frequency data. A large unsigned likelihood-ratio value signals departure from independence; the observed and expected counts must still be compared to determine its direction.

Association can also be directional. The conditional probabilities P(y∣x)P(y\mid x) and P(x∣y)P(x\mid y) need not be similar. An uncommon adjective may strongly predict a noun, while that noun, which occurs with many adjectives, may only weakly predict the adjective. Measures such as directional differences in conditional probability address this asymmetry. They answer a different question from a symmetric measure such as PMI: how informative is encountering one item for predicting the other?

No ranking measure defines linguistic conventionality by itself. High association may reflect a proper name, a news event, a quotation, or a conventional phrase. Low association may occur for an important combination containing a very frequent word. Statistical significance is also distinct from practical or linguistic importance: sufficiently large corpora can make small departures from independence statistically conspicuous. Interpretation therefore depends on frequency, effect size, dispersion, grammatical relationship, and inspection of actual examples, rather than on a single numerical threshold.

Meaning, discourse, and variation

Collocation contributes to lexical semantics by showing how meanings are distributed across contexts. The adjective strong behaves differently in strong coffee, strong evidence, and strong opposition. These uses share a broad association with intensity or effectiveness, but their precise interpretation depends on the noun and the conventional expression. Collocational evidence can support the identification of word senses, reveal metaphorical extensions, and show where dictionary distinctions correspond to different patterns of use.

A related concept is semantic preference: a word’s tendency to occur with members of a semantic class rather than merely with individual words. A verb may favor expressions referring to quantities, difficulties, or institutional actions. Semantic prosody concerns the evaluative or discourse orientation associated with recurrent surroundings. Research by Sinclair and, prominently in the 1990s, Bill Louw drew attention to expressions whose typical environments convey favorable, unfavorable, or otherwise attitudinal meanings not obvious from an isolated definition.

The often-discussed verb cause, for example, frequently appears with nouns denoting undesirable events or conditions, including damage, problems, and death. This distribution does not make cause intrinsically negative or prevent combinations such as cause delight. It shows that a word’s recurrent environments can create expectations extending beyond its basic denotation. Analyses must distinguish the evaluative character of individual collocates, the larger phrase, and the discourse act being performed. The scope and interpretation of semantic prosody remain subjects of theoretical discussion.

Variation across registers is equally important. Academic writing, legal documents, conversation, advertising, and sports commentary develop characteristic combinations because they repeatedly perform different communicative tasks. Academic prose often uses patterns such as statistically significant difference and the results suggest, while institutional texts contain recurrent procedural wording. Regional differences also occur: make a decision and take a decision have different distributions across varieties and contexts of English. Neither a dictionary label nor a frequency comparison should erase the role of genre within those regional patterns.

Collocations can change as practices, technologies, and social institutions change. Expressions such as cloud computing and social distancing became prominent through particular technological or public contexts, though the histories of the expressions are not reducible to their periods of greatest visibility. Repeated metaphorical combinations may become ordinary expressions, linking collocation with metaphor and semantic change. Creative writing and humor can exploit expectations by replacing a familiar collocate with an unexpected one. Such deviations work precisely because conventional associations are recognizable; they are not evidence that every unusual combination constitutes an error.

Acquisition, processing, and language teaching

Collocational knowledge is an important component of language acquisition. Children encounter words in recurring utterances, not as entries in an isolated vocabulary list. Repeated sequences supply evidence about word boundaries, grammatical patterns, and communicative routines. Early formulas and partially analyzed expressions can later support more flexible production. This does not mean that every familiar phrase is stored as an indivisible unit; learners can retain information about a sequence while also analyzing its internal structure.

In second-language acquisition, knowledge of individual words often develops faster than command of their conventional combinations. A learner may understand decision, make, and do yet produce do a decision under the influence of a broader verb category or another language. Collocation thus helps explain the difference between being intelligible and producing wording characteristic of a particular speech community. Difficulties may involve selecting the lexical partner, choosing the appropriate grammatical frame, or recognizing a restriction that has no close equivalent in the learner’s other language.

Psycholinguistic studies examine whether familiar sequences are processed more efficiently than matched unfamiliar combinations. Methods include lexical decision, self-paced reading, eye tracking, and electrophysiological recording. Frequency, predictability, semantic transparency, and participant proficiency can all affect the results. A processing advantage for a frequent phrase does not by itself prove that it has a single dedicated mental representation. It may arise from strengthened associations among words, constructional knowledge, contextual prediction, or several mechanisms acting together.

Teaching approaches commonly make collocation visible through contextual examples, collocation dictionaries, grouped lexical patterns, and corpus concordances. A lesson centered on research, for instance, can distinguish conduct research, carry out research, research findings, and research into a topic, while explaining their grammatical differences. Receptive recognition and productive command are separate outcomes: recognizing draw a conclusion in a text is less demanding than retrieving it during speech. Tasks and assessments therefore produce different pictures of a learner’s knowledge.

Frequency alone is an insufficient basis for selecting teaching material. A moderately frequent expression may be essential in a learner’s professional domain, while a very common sequence may already be transparent and familiar. Measures of collocational knowledge also require care: an exercise allowing only one answer can incorrectly reject a legitimate alternative, and judgments may vary by register or region. Language-learning research accordingly distinguishes conventionality, appropriateness, fluency, and grammatical accuracy rather than treating all departures from a selected expression as the same kind of deficiency.

Dictionaries, terminology, and translation

Dictionaries represent collocation because definitions alone do not show how a word combines with others. A learner’s dictionary may explain the meaning of attention while separately illustrating pay attention, attract attention, and close attention. Specialized collocation dictionaries organize such information around bases and grammatical patterns, enabling a user to find verbs, adjectives, and other partners associated with a target word. The BBI Combinatory Dictionary of English, first published in 1986, and the Oxford Collocations Dictionary for Students of English, first published in 2002, are prominent examples of this reference tradition.

Corpus-based lexicography adds distributional evidence to editorial judgment. Frequent combinations can guide examples, sense distinctions, grammatical codes, and usage labels. Nevertheless, a list of statistically associated words is not yet a dictionary entry. Editors must separate senses, distinguish general combinations from domain-specific ones, identify genuine syntactic relationships, and decide which patterns help a particular readership. An unusual but established technical collocation may deserve inclusion even when it has low frequency in a general corpus.

In terminology work, collocations are related to but distinct from terms. Acute myocardial infarction names a medical condition and functions as a term, whereas verbs characteristically used with a term may describe diagnosis, measurement, manufacture, or treatment without themselves becoming part of the designation. A technical translator often needs both kinds of knowledge: the accepted name of an object or process and the conventional wording used to describe its behavior. Termbases that record only isolated equivalents may therefore leave important production problems unresolved.

Collocation is a persistent issue in translation studies because lexical equivalence does not guarantee combinational equivalence. English make a decision corresponds naturally to French prendre une décision, whose verb ordinarily means “take.” English strong tea can correspond to French thé fort, but similar-looking adjective correspondences do not extend uniformly across all nouns. Translating each component independently can produce a grammatically interpretable sequence that does not match established target-language usage. Translation equivalence must therefore be assessed at the level of expressions and communicative context as well as individual words.

A parallel corpus can reveal how expressions are rendered in translated texts, while comparable corpora can show how the same subject is discussed independently in different languages. Neither provides a mechanical solution: translations may reflect source-language influence, and comparable texts may differ in genre or institutional setting. Word alignment can also fail when one language expresses a relation through a verb and another through a noun phrase or inflected form. Translators consequently evaluate collocations together with meaning, register, terminology, and the function of the whole passage.

Computational extraction and language technology

Automatic collocation extraction typically combines linguistic preprocessing with statistical ranking. A system may tokenize text, identify lemmas and parts of speech, collect candidate pairs or phrases, and rank them using association measures. Candidate selection can rely on surface windows, recurring sequences, or grammatical dependencies. Each design captures different phenomena: adjacent sequences are straightforward to count, while dependency relations can connect words separated by modifiers or reordered through syntactic operations.

Collocations overlap with multiword expressions, a broader computational category including idioms, compound-like units, conventional phrases, and constructions with lexical restrictions. Detecting a frequent sequence is not equivalent to identifying a multiword expression. A sequence may be frequent because it contains common grammatical words, while an established expression may be discontinuous or rare. Systems therefore combine frequency with syntactic patterns, measures of lexical substitutability, semantic information, and annotated examples. Evaluation depends on the intended category and application rather than on a universal inventory.

Statistical language modeling incorporates collocational information through conditional prediction. An n-gram language model estimates a word’s probability from a limited sequence of preceding words, directly capturing many local combinations. Broader distributional models represent words through their contexts. Methods such as GloVe use word co-occurrence statistics to learn vector representations, but similarity in such a representation should not be confused with collocation: two near-synonyms may have similar contextual distributions without frequently occurring beside one another.

A large language model can represent much longer contextual dependencies and generate conventional combinations without consulting an explicit collocation dictionary. Its output nevertheless reflects the composition and distribution of its training data, and fluent wording is not proof of factual accuracy or domain appropriateness. Specialized combinations, recent terminology, and less-represented language varieties may require independent evaluation. Explicit collocation resources remain useful when an application needs interpretable evidence, stable terminology, or a clear explanation of why one expression is preferred.

Practical applications include writing assistance, query expansion, terminology extraction, word-sense disambiguation, and machine translation. A writing assistant may identify an unusual verb–noun combination and offer alternatives attested in comparable contexts. A search system may recognize a multiword expression as a unit rather than treat its components independently. Performance is assessed against annotated examples, corpus evidence, and task-specific outcomes, with particular attention to false corrections: a system that replaces every uncommon combination with a common one can erase technical distinctions, regional usage, or deliberate stylistic innovation.