The 385+ Million Word Corpus of Contemporary American English (1990-present): A new tool for examining language variation and change

The last 15-20 years have seen the introduction of
“mega-corpora” such as the Bank of English and
the British National Corpus, which contain anywhere
from 100-500 million words. Until recently, however,
there have been no large, balanced corpora of American
English. Only a small portion of the Bank of English,
for example, is from the US. The well-known American
National Corpus has not been updated in several years,
it has only 22 million words of text, and it is quite unbalanced
in terms of genre representation (essentially no
fiction, no popular magazines, etc). On the other hand,
there are other large “corpora” of American English
(such as the GigaWord collection of newspaper articles),
but these represent just one genre.
The situation has changed recently, with the recent introduction
of the Corpus of Contemporary American
English (COCA) (www.americancorpus.org), which we
released in Spring 2008. This is the first large, balanced
corpus of American English, and it will permit researchers
to address many questions related to language change
and linguistics variation, which could not have been answered
until this time. The corpus is composed of more
than 385 million words of text in more than 150,000 articles
and books, with at least 20 million words each year
from 1990 to 2008 (and it will be updated from this point
on as well). In each year, the corpus is divided into five
equally-sized genres: spoken (transcripts of unscripted
conversation on 100+ TV and radio programs each year),
fiction (novels, short stories, and movies scripts), popular
magazines (100+ magazines each year), newspapers
(ten newspapers from across the US), and academic
journals (100+ journals each year). The wide range of
genres means that researchers can study in detail variation
between these genres, and the consistency in genres
across time means that researchers can accurately study
linguistic changes. In addition, the corpus is tagged and
lemmatized (using CLAWS, the same tagger that was
used for the British National Corpus), which greatly facilitates
syntax-oriented queries. The entire corpus architecture and interface are designed
to facilitate research into language variation and change.
Users can quickly and easily find the frequency of any
word, phrase, substring (e.g. suffixes), or syntactic construction
in each year since 1990, and in each of the
five major registers. Example might be words such as
carbon-neutral or the quotative like, phrases like perfect
storm or tipping point, suffixes like –gate (Iraqgate or
zippergate), or grammatical constructions like preposition
stranding, zero relative clauses, or the ‘get passive’.
They can also see detailed information on frequency and
distribution of words and constructions in micro-genres,
such as the rise of bling in African-American and entertainment-
related popular magazines.
The corpus also allows users to find the collocates in different
genres and groups of years since 1990, which can
provide valuable insight into semantic change and variation.
For example, they can compare the collocates of
woman or of peace in spoken, fiction, and newspapers
to see how these concepts are viewed and discussed differently
in the two genres. They can also compare collocates
over time, such as the increasingly positive collocates
with geek since the early 1990s, or the increasing
environmental emphasis over time, evidenced by new
collocates with green. Finally, they can easily compare
the collocates of two words to see contrasts in the meaning
or usage of the two terms. Examples might be adjectives
with Democrats vs. Republicans (electable, fun,
open-minded vs. extremist, mean-spirited, and greedy),
or bias in the collocates with women and men (glamorous,
real-life, disadvantaged vs. honorable, self-made,
and wise).
Other features allow for fairly complex semanticallyoriented
searches. Due to the relational database architecture,
we have been able to integrate a thesaurus with
entries for 60,000+ headwords, as well as WordNet,
and users can also create “customized lists” on the fly.
These allow for rather powerful queries, such as “any
form of any synonym related to the verb clean within
five words of any word in the ‘houseItems’ list created
by Jones” (clean the pots, washing some windows, the
floor he mopped) or “any hyponym of emotions within
five words of a word in the ‘familyTerms’ list created by
Smith (Grandpa seemed to be pretty happy, the excited
children, the moms that are most worried). As can be
seen, this goes far beyond most other corpus architectures,
which are often limited to just word, phrase, lemmas,
or parts of speech.
This example of semantically-oriented searches leads us
finally to a brief discussion of the overall challenge of
designing an architecture that achieves the three competing
goals of 1) size 2) speed, and 3) annotation. Achieving
two of three goals is relatively simple, but achieving
all three simultaneously – in the real world – is much
more difficult. For example, there are many search engines
that allow fast retrieval from very large “corpora”
or text archives (e.g. Google or Lucene), but which allow
for little if any annotation (e.g. even basic part of
speech tagging or lemmatization, much less integration
with thesauruses or user-defined lists). Other approaches
provide speed and annotation, but are completely inadequate
in terms of scalability – either in terms of size
and/or speed. There is no limit to the number of proprietary
architectures that have been designed over the past
decade or two, and which might work very well for a
small one million word corpus, but which are utterly unscalable.
A query might take just two or three seconds
for a well-annotated 10 million word corpus, but (assuming
linearity), that same query then takes two minutes or
more for a 350-400 million word corpus.
Our approach – which is based on a (still) proprietary architecture
involving relational databases and a massively-
redundant n-grams schema – is one of just a handful
that adequately allows for size, speed, and annotation.
Even a complicated query– involving part of speech,
lemma, synonyms, customized word lists, and limited
by sub-genres – typically takes just 2-3 seconds to generate
results from the entire 385+ million word corpus.
In addition, ours is the only architecture (as far as we are
aware) that allows for such a wide range of comparisons
– e.g. across sub-corpora, or the collocates of different
words. For example, SketchEngine allows comparisons
between different words, but not by sub-corpus. The IMS
Corpus Workbench (CWB) allows comparisons between
sub-corpora, but not between different words. Ours offers
both of these, full integration with thesauruses and
lexical resources like WordNet, as well as much more.
In summary, the Corpus of Contemporary American
English (COCA) is based on an architecture and interface
that allows for a wide range of queries, and which
does so quickly and easily. In terms of the textual database,
it is both large (385+ million words, and growing)
and well-balanced (in terms of genres and sources).
All of these features serve to create a unique resource
that allows researchers to look at a wide range of questions
dealing with recent changes and current variation
in American English, which would have been difficult or
impossible to investigate before this time.