I've noticed that Ngrams feature quite often in threads I've been involved with (I've used it myself) but, allowing that the Ngram program can be useful, how cautious should we be in using it?

An extreme example is the phrase "fit down". This phrase was enormously popular in 1800, followed by a precipitous decline through 1820, virtually vanishing as we move from 1820-1840.

The "fit down" Ngram

enter image description here

Unfortunately, the tale of "fit down" is entirely the result of the google OCR program insisting in transcribing the "long-s" as "f". What we are seeing is the ghost of "sit down".

How wary should we be of the Ngram program? In just what ways is it useful? Should there be lodged, somewhere, a "Guide to Using Ngram"?

  • 2
    While the Ngram system is a valuable tool, it needs to be used with care. In particular there are numerous errors both in OCR encoding and in the recording of the dates of sources. One should not trust the result to any great degree unless a sample of the "hits" have been examined directly to verify their correctness. And, unfortunately, many of the sources have copyright restrictions that don't allow direct viewing of the source material, leaving you kind of stuck.
    – Hot Licks
    Jun 9, 2017 at 0:56
  • 2
    (Another "gotcha" is that Ngram ignores all punctuation, and it may find your search phrase as the end of one line and the start of another, creating a bogus "hit".)
    – Hot Licks
    Jun 9, 2017 at 1:05
  • 5
    The usefulness and limitations of Ngrams topic has certainly been discussed on Meta before now. // Your bogus 'fit down' example is a good one. Individual results may be examined to estimate how many false positives might be expected, and related searches ('fitting on his chair'?) compared to spot anomalies. That said, Ngram results are usually better evidence than 'My mates and I say it this way' / 'As far as I'm concerned, nobody uses this'. // This topic merits a doctoral thesis. Jun 9, 2017 at 1:32
  • 2
    Using corpus data is much much better than not using corpus data, but you do have to be aware of the limitations of the corpus. If you post Google NGrams then you have to consider whether the data you're using could be the result of bad OCRing. Jun 9, 2017 at 1:38
  • I'd agree that a decent corpus is desirable, but Ngram is doubly flawed -- it depends on the corpus contained in google books (which is incomplete and biased, and may draw on duplicate copies to generate false figures -- how many editions of Jonathan Swift are being digested and shoveled into Ngram?), added to which is the entirely mechanical (and not terribly sophisticated) google OCR program. For any kind of detail, it's absolutely necessary to go to google books, so why not just search there in the first place? Jun 9, 2017 at 1:56
  • I'm also more than slightly suspicious of corpus linguistics when it uses the corpus as an end point rather than a starting point -- collect the material, tag it, and churn out the numbers. That said, I use the OED and GDoS more as corpora than as dictionaries. The Old Bailey Online (though not intended as such) provides a powerful set of texts from 1674-1913, and can be at times more useful for exploring the nuances of terms, and when they appear and disappear, than any other source I've found. Then there are the other dictionaries -- LEME, SLD, MED. Have I missed anything? Jun 9, 2017 at 2:07
  • 1
  • Here's an article on this topic: wired.com/2015/10/…
    – Xanne
    Jun 9, 2017 at 9:49
  • There's certainly an overlap -- I noticed that the point I made here [above] about the problems of multiple editions skewing the results comes up there too. In this area, as to Swift, "sodder and patch up" / "solder and patch up" produces "interesting" results. Would a fair summary be, "Ngram is good for what it's good for, and bad for everything else."? Jun 9, 2017 at 12:48
  • 1
    @curiousdannii 1- I'm having a hard time thinking that this is not a duplicate. 2- That said, there's nothing that says "Here are the ways that Google NGrams should be used properly 3- revisiting problems like this is useful for re-evaluating the problems (GNG has added some search features since it first came out) and for re-educating us all about drawbacks and mitigations.
    – Mitch
    Jun 9, 2017 at 13:04


Browse other questions tagged .