English-Corpora.org

The Corpus of Historical American English and small corpora (Brown+, ARCHER, etc)

There are a number of other corpora of historical English for the 1800s-1900s. Some of the better-known ones are the BROWN family of corpora, ARCHER, CONCE, and DCPSE. The data from these corpora has resulted in many insightful studies.

Like these other corpora, COHA is well-balanced in terms of genres and the corpus has roughly the same genre balance from decade to decade. As a result, it is not surprising that -- for the constructions where there is enough data in small corpora -- the COHA data is quite similar. This should be reassuring to those who are used to the more "established" corpora.

To give a few examples, the following table shows data from COHA, TIME, and COCA for some constructions that have been researched in recent studies using the BROWN family of corpora (hereafter BROWN+; 4 million words 1961 / 1991; Brown, LOB, FROWN, FLOB)

Note that -- in order to simplify this web page -- these links are just for one-step queries. For some of the phenomena, we'd need to run more than one query and adjust the frequencies. For example, with that/which as relative pronouns (the argument that/which he makes), we would run a second query (using customized wordlists) to find noun complements (e.g. the fact that they won't be here) and then subtract those from the relative pronoun counts.

There are two important differences between COHA and these other corpora, however. The first is that COHA has an architecture and interface that allow researchers to look at many kinds of  phenomena that would be difficult or impossible to study otherwise -- in terms of morphological, syntactic, semantic, and lexical change.

The second main difference between COHA and corpora like ARCHER, CONCE, DCPSE, and the BROWN family of corpora relates to size. COHA is about 100-475 times as large as the four corpora listed above. In addition, the COHA texts provide data that are continuous, meaning that they sample the language every single year from 1810-2009, rather than just every 30 years or so. (For example, there are about 2 million words each year from the 1880s-2000s). Because of its size and continuous nature, COHA provides robust, granular data that is impossible with the other corpora.

Because there is so much continuous data, we can look at specific changes in incredible detail, and then compare those changes to others that are occurring at about the same time, to see how the changes are related. Let's look at a quick example. The following chart shows the shift from to-V (he started to sing) to V-ing (he started singing) with a number of verbs during the past 200 years.

Even though the chart has a lot of lines, notice how the first significant shift towards V-ing seems to occur with start in about 1900-1920s, followed by the less frequent but semantically-related verb begin (the red lines). Then in about the mid-1900s, there is an increase with the related "emotion" verbs like, love, and hate (the green lines), with the biggest increases with the emotionally strongest verbs -- love and hate.

Here's the point, though. The following is the data for hate (e.g. I hate to write papers > I hate writing papers).

Imagine that instead of 475 million words (the size for COHA), we had a corpus 1/100th or 1/200th that size -- or in other words, the size of ARCHER, CONCE, the BROWN family, etc. Rather than 300-400 tokens in a given cell in the table above, we'd have 1 or 2. With such sparse data, we couldn't really map out the shifts with any given verb or see the relationship between the different verbs.

The example above deals with syntactic change. We could repeat this example with any number of other examples in syntax or in other areas dealing with language change. Here's just a few:

• (lexical) verbs with up in the 1880s-1920s (left) compared to the 1960s-2000s (right)

• (morphological) -able adjectives 1810s-1910s (left) compared to 1920s-2000s (right)

• (semantic) collocates of gay in the 1840s-1910s (left) compared to the 1970s-2000s (right)

In each case, the number of tokens with a given word or collocate occurs just 20-60 times, even in the 475 million word corpus. In a small 2-4 million word corpus, it would occur at about 1/100th or 1/200th this rate -- or in other words, maybe one or two tokens. That would not be enough to look at any of these -- or any similar -- changes.

With a small 2-4 million word corpus, we are limited to looking at just high frequency phenomena -- like modals, passives, perfects, progressives, prepositions, conjunctions, and relative pronouns. There has been some great research done on these topics over the years, by some of the best researchers in the field of Late Modern English. But after 20-30 years of research on this handful of phenomena, we would suggest that it's time to move on to a wider range of phenomena. The 475 million word Corpus of Historical American English is arguably the only publicly-available, structured corpus of historical English that allows us to do so.