Corpus Linguistics for Language Learning

Corpus linguistics is the study of language through large, principled collections of real texts — called corpora — analysed with the help of computers. Instead of relying on a linguist's intuition about what is possible in a language, it asks a different question: out of everything speakers and writers could say, what do they actually say, and how often? That shift — from invented examples to observed evidence — quietly changed how dictionaries are written, how coursebooks choose their words, and how we think about what it means to know a language. For anyone learning a foreign language, the payoff is concrete: corpora tell us which words matter most, and how words really combine in the wild.

The field rests on a simple idea with far-reaching consequences. Native intuition is excellent at judging whether a sentence is possible, but poor at judging what is typical. Ask a fluent speaker how often they use a given word, or which words usually accompany it, and they will guess — and often guess wrong. A corpus does not guess. It records millions of words of genuine usage and lets patterns emerge from the data.

What a corpus is: from the Brown Corpus to the billion-word web

A corpus is a large, structured sample of language, assembled to be representative of a variety — a language, a dialect, a register — and stored so a computer can search it. The modern field begins in 1961, when W. Nelson Francis and Henry Kučera at Brown University compiled the Brown Corpus: one million words of American English, drawn in equal 2,000-word samples from 500 texts across 15 genres, from press reportage to fiction. By today's standards a million words is tiny, but the Brown Corpus was the first machine-readable corpus built for linguistic research, and it set the template — balanced sampling, documented sources, computable format — that everything after it followed. Its British twin, the Lancaster–Oslo/Bergen (LOB) Corpus, deliberately mirrored its design so the two varieties could be compared directly.

Corpora then grew by orders of magnitude. The British National Corpus (BNC), completed in the mid-1990s, holds 100 million words of written and spoken British English. The Corpus of Contemporary American English (COCA), compiled by Mark Davies from 2008 onwards, began at over 385 million words and has since passed a billion, balanced across spoken language, fiction, magazines, newspapers and academic prose. Alongside these, learner corpora (collections of texts written by language learners) and multilingual corpora opened up new kinds of study. Bigger corpora matter because language is deeply patterned but also sparse: many important word combinations appear only a few times per million words, so you need a great deal of text before the pattern becomes visible.

What do you actually do with a corpus? Three basic operations underlie almost everything:

Concordances. A concordance line shows a search word centred in its immediate context — the "keyword in context" view — with every occurrence stacked one above the next. Reading down the column, patterns jump out that no single example would reveal: which prepositions follow a verb, whether a word is usually positive or negative, what typically comes just before it.
Frequency. The corpus counts how often each word and structure occurs, producing frequency lists that rank vocabulary from the most common to the rarest — the raw material for deciding what to teach first.
Collocation. The corpus reveals which words habitually keep company with which others — strong tea but powerful engine, make a decision but do your homework. These preferences are largely arbitrary, invisible to grammar, and precisely what learners most often get subtly wrong.

Collocation, frequency and the COBUILD revolution

The linguist J. R. Firth captured the central insight in a famous line: "You shall know a word by the company it keeps." A word's meaning and behaviour are bound up with the other words it appears alongside — and only a corpus can show that company at scale. Nobody pursued this further than John Sinclair, who from the 1980s led the COBUILD project at the University of Birmingham, building the corpus later known as the Bank of English specifically to study how English words really behave.

From that data Sinclair drew a radical conclusion, which he called the idiom principle: much of language is not assembled word by word according to grammar rules, but reused in ready-made chunks — semi-fixed phrases and habitual collocations that speakers select whole. We say "to a certain extent", "as a matter of fact", "I'm afraid I can't" not by free grammatical construction but as prefabricated units. The corpus showed that this pre-assembled, formulaic layer is not a marginal curiosity but a huge part of ordinary language — an idea explored in our entry on formulaic sequences.

This had an immediate practical result. The Collins COBUILD English Language Dictionary (1987) was the first learner's dictionary built entirely from corpus evidence: its headwords were chosen by frequency, its definitions described how words were actually used, and — crucially — its example sentences were taken from real texts rather than invented by lexicographers. Every major learner's dictionary and much grammar reference has followed corpus methods ever since. When a modern dictionary tells you a word is "formal" or "mostly used in the negative", that judgement now typically rests on counted evidence, not on an editor's hunch.

Data-driven learning: the learner as researcher

If corpora could correct the intuitions of dictionary editors, why not put them directly into the hands of learners? That was the proposal of Tim Johns, also at Birmingham, who in 1991 coined the term data-driven learning (DDL). His slogan was to "cut out the middleman as far as possible and give the learner direct access to the data." Instead of being handed a rule about, say, when to use make versus do, the learner is given a set of concordance lines and asked to work the rule out — to become, in his phrase, a language detective.

The reasoning is that a learner who discovers a pattern from real evidence understands it more deeply, and remembers it better, than one who is simply told. It reframes the classroom: the teacher stops being the sole authority on the language and becomes a guide to the evidence; the learner does the noticing. Johns called the practice "classroom concordancing", and it remains influential — behind today's online corpora, concordancers and the everyday habit of checking how a phrase is really used before committing to it. It also connects naturally to the broader case that understanding language you can mostly follow drives acquisition, discussed in our entry on comprehensible input.

What the learner takes from it: frequency lists and authentic collocations

Two corpus products reach learners most directly. The first is the frequency-based word list. Michael West's General Service List (1953) was an early, hand-counted attempt to identify the few thousand words that give the most reading coverage. Its modern successor, the New General Service List (Browne, Culligan and Phillips, 2013), was derived from a 273-million-word section of the Cambridge English Corpus: its roughly 2,800 words provide about 92% coverage of general English text. The lesson for a learner is bracing — a relatively small, well-chosen core of vocabulary unlocks most of everyday language, and frequency data tells you exactly which words that core contains. Time spent on the top few thousand words pays off far more than time spent on rare ones.

The second product is knowledge of authentic collocation — how words genuinely combine, as opposed to what a plausible-looking translation suggests. This is where intuition, including a teacher's, most often fails, and where corpora most often surprise. Corpus studies have repeatedly shown that coursebooks over-teach some structures and under-teach others, and that the invented example sentences of older textbooks frequently do not match how the words behave in real use. A learner who checks a phrase against a corpus — or uses materials built on one — is learning the language as it is actually spoken and written, not an idealised textbook version of it.

What this means for language learning

The practical message of corpus linguistics for a learner is to trust evidence over intuition, and to learn language in the units that real usage comes in. Prioritise the high-frequency core: the few thousand most common words and their most common collocations do the overwhelming majority of the work, and frequency lists tell you what they are. Learn words together with their habitual company rather than in isolation, because "strong tea" and "make a decision" are facts about usage that no rule will give you. And treat whole, authentic sentences as the natural unit — a sentence carries a word's real collocations and grammar together, in context, which is exactly why learning in full sentences operationalises the corpus insight so directly. This is the thinking behind the Taalhammer method: practising real sentences, chosen for usefulness, so that frequent words and their authentic combinations are met and rehearsed the way corpora show the language actually works.

Frequently asked questions

What is a corpus in simple terms?

A corpus is a large, organised collection of real texts — spoken and written — stored so a computer can search it. Its purpose is to show how a language is actually used rather than how we imagine it is used. The first machine-readable one was the Brown Corpus (1961) with a million words; modern corpora such as COCA contain a billion or more.

What is a concordance and why is it useful?

A concordance lists every occurrence of a search word with the text on either side of it, aligned so the word sits in a central column. Reading down that column, you see the patterns a single example would hide: the prepositions a verb takes, the words that habitually accompany it, whether it tends to be positive or negative. It is the main tool through which corpus evidence becomes visible.

How does corpus linguistics help language learners?

In two main ways. Frequency data identifies the high-value vocabulary to learn first — a few thousand words cover most everyday language. And collocation data shows how words really combine, correcting the intuitions that lead learners into unnatural phrasing. Corpus evidence also underlies the dictionaries, coursebooks and word lists learners already rely on.

Sources

W. Nelson Francis, Henry Kučera, Computational Analysis of Present-Day American English, Brown University Press, 1967 (Brown Corpus, 1961).
J. R. Firth, "A Synopsis of Linguistic Theory, 1930–1955", in Studies in Linguistic Analysis, Blackwell, 1957.
John Sinclair, Corpus, Concordance, Collocation, Oxford University Press, 1991.
John Sinclair (ed.), Collins COBUILD English Language Dictionary, Collins, 1987.
Tim Johns, "From Printout to Handout: Grammar and Vocabulary Teaching in the Context of Data-Driven Learning", in T. Odlin (ed.), Perspectives on Pedagogical Grammar, 1994; and T. Johns, P. King (eds.), Classroom Concordancing, ELR Journal 4, 1991.
Michael West, A General Service List of English Words, Longman, 1953.
Charles Browne, Brent Culligan, Joseph Phillips, The New General Service List, 2013 (newgeneralservicelist.org).
Mark Davies, "The 385+ million word Corpus of Contemporary American English (1990–2008+): Design, architecture, and linguistic insights", International Journal of Corpus Linguistics 14(2), 2009.