Written and maintained by CASRAI Editorial Board
Last updated
The Corpus of Contemporary American English (COCA) is a roughly one-billion-word collection of American English texts from 1990 to 2019, balanced across eight genres, that linguists use to count how often words and phrases occur. This guide gives its size and composition from the publisher’s own documentation, shows how to turn a raw hit count into a frequency per million words, and explains how to cite it.
What is the Corpus of Contemporary American English?
COCA is a text corpus: a large, structured sample of real language held in a database so that words, collocations and grammatical patterns can be searched and counted. It is part of the English-Corpora.org family of corpora created by Mark Davies. According to the English-Corpora.org documentation it contains 485,179 texts and 1,002,889,754 words, covering 1990 to 2019, and the site states that the genre balance is essentially the same in each of those 30 years. The Language Bank of Finland, which distributes a licensed research copy, lists the same scale (about 1 billion words, 485,000 texts, March 2020 release). Corpus work of this kind belongs to linguistics more broadly, and especially to the study of usage and variation described in sociolinguistics.
COCA composition by genre
COCA is divided into eight genres of roughly equal size. The text and word counts below are from English-Corpora.org’s representativity document (read 2026-11-12); the share column is our own calculation (genre words divided by the corpus total).
| Genre | Texts | Words | Share of corpus |
|---|---|---|---|
| TV/Movies | 23,975 | 129,293,467 | 12.9% |
| Spoken | 44,803 | 127,396,932 | 12.7% |
| Fiction | 25,992 | 119,505,305 | 11.9% |
| Magazine | 86,292 | 127,352,030 | 12.7% |
| Newspaper | 90,243 | 122,958,016 | 12.3% |
| Academic | 26,137 | 120,988,361 | 12.1% |
| Web/Blog | 98,748 | 125,496,216 | 12.5% |
| Web/General | 88,989 | 129,899,427 | 13.0% |
| Total | 485,179 | 1,002,889,754 | 100% |
The eight word counts sum exactly to the stated total of 1,002,889,754. Each genre holds between 11.9% (Fiction) and 13.0% (Web/General) of the words. Genres differ widely in how many texts supply those words: Web/Blog has 98,748 texts, while TV/Movies has 23,975, so the average text is much shorter in some genres than others. The Language Bank of Finland README describes the same eight categories (spoken, fiction, magazine, newspaper, academic, blogs, web pages, TV and movie subtitles) at about 125 million words each; blogs, web pages and subtitles were added in the 2020 update.
Worked example: frequency per million words
Raw counts from different corpora are not comparable, because the corpora differ in size. The standard fix is to normalise to a frequency per million words: divide the hit count by the corpus size in millions.
COCA size in millions: 1,002,889,754 / 1,000,000 = 1,002.889754 million words.
The English-Corpora.org document compares COCA with the British National Corpus (BNC, 1994 texts), which it gives as 4,053 texts and 97,619,311 words, about one tenth of COCA’s size. BNC size in millions: 97.619311.
Hypothetical illustration. Suppose a search returned 1,250 hits for the same word in each corpus. These counts are invented for the arithmetic only and are not real results.
- COCA: 1,250 / 1,002.889754 = 1.25 per million words (1.246 before rounding).
- BNC: 1,250 / 97.619311 = 12.81 per million words (12.805 before rounding).
Identical raw counts therefore imply a ten-fold difference in relative frequency, which is the whole reason to normalise. The same logic applies within COCA: to compare a word across genres, divide each genre’s hits by that genre’s word count, not by the whole corpus.
Interpreting COCA counts
| What you see | How to read it |
|---|---|
| Raw hit count | Occurrences in the whole corpus or a chosen section; meaningful only alongside the size of that section. |
| Frequency per million | Hit count divided by section size in millions; the figure to quote when comparing sections or corpora. |
| Genre difference | Compare per-million rates across the eight genres, whose sizes are 119.5 to 129.9 million words each. |
| Texts versus words | A term concentrated in few texts is less widespread than the same total spread across many; check text counts, not only word counts. |
Limits and common mistakes
- Comparing raw counts across corpora of different size. Normalise first, as above.
- Treating the corpus as all of English. COCA covers American English from 1990 to 2019 only; it does not sample British, other national varieties or earlier periods.
- Assuming perfect balance. The balance across genres is the publisher’s design claim. English-Corpora.org defends the corpus’s representativity in its own document, and a published review by Egbert, Biber and Gray (2022) is mentioned there; treat balance as an assumption to report, not a proven property.
- Ignoring genre mix. Because web genres, spoken material and subtitles are included alongside edited prose, a pattern’s frequency can reflect genre rather than general usage.
- Expecting complete text in licensed downloads. The Language Bank of Finland copy replaces 10 words in every 200 with the symbol @ to stay within US fair-use rules, which affects any analysis of running text or longer sequences.
Access and licensing
COCA is searched through English-Corpora.org. Licensed copies for research exist: the Language Bank of Finland distributes COCA (VRT format, 2020) under a CLARIN RES licence, and the University of Virginia Library lists a licensed copy limited to University of Virginia research use, with full-text redistribution restricted. Check the licence of whichever copy you use before sharing any text from it. For a contrasting web-scale text resource, see the CASRAI guide to Common Crawl.
How to cite COCA
The University of Virginia Library entry says publications using the corpus must cite it and points to corpusdata.org. A typical reference names the corpus, its creator and the access address, for example (APA style, adapt to your edition and the version you used):
Davies, M. (2008-). The Corpus of Contemporary American English (COCA). https://www.english-corpora.org/coca/
State in your methods section the date you ran queries, the genres and years included, and whether you used raw counts or per-million rates.
Frequently asked questions
What is COCA?
COCA is a billion-word corpus of American English texts from 1990 to 2019 used to study word and phrase frequency, collocation and grammar.
How big is the Corpus of Contemporary American English?
1,002,889,754 words in 485,179 texts, according to English-Corpora.org.
What years and genres does COCA cover?
1990 to 2019, across eight genres: TV/Movies, Spoken, Fiction, Magazine, Newspaper, Academic, Web/Blog and Web/General.
How does COCA differ from the BNC?
The BNC (1994) has 4,053 texts and 97,619,311 words, about one tenth of COCA, and is British English from a single period; COCA is American English spanning 30 years.
How do I compare a word’s frequency in COCA with another corpus?
Divide each hit count by that corpus’s size in millions of words and compare the per-million rates.
Sources
- English-Corpora.org, Representativity (genres): https://www.English-Corpora.org/help/representativity.pdf (read 2026-11-12).
- Language Bank of Finland (Kielipankki), COCA VRT 2020 README: https://www.kielipankki.fi/download/coca/coca-vrt-2020/README.txt (read 2026-11-12).
- University of Virginia Library, Corpus of Contemporary American English (COCA) data source entry: https://library.virginia.edu/data/datasources/licensed/corpus-of-contemporary-american-english-coca (read 2026-11-12).
