Archived Science Blog page from 2007. It stays on Science Blog so the original link still works. It is not a current article. Back to Science Blog
GoogleScience
A Google search can help you find cutting-edge research. A Google search can also
Many questions in linguistics (the formal study of language) and psycholinguistics (
) are answered by turning to a corpus. A corpus is a large selection of texts and/or transcripts. Just a few years ago, they were difficult and expensive to create. Arguably the most popular word frequency corpus in English -- the verable
-- was based off of one million words of text. One million words sounds like a lot, but so many of those are "the" and "of" that in fact many words do not appear in the corpus at all.
So what does one do with a corpus? One obvious thing is to figure out which words are more common than others. The most common words in English are short function words like "a" (found on 4.95 billion web pages) and "of" (3.61 billion pages). Google, of course, doesn't tell you how many times a words appears, but only on how many pages it appears...which may actually be an advantage, since we're typically more interested in words that show up on many websites than a word that shows up many, many times on just one website.
You might be interested in what the most common noun is. Is it "time," "man," "city," "boy," or "Internet?" You might check to see whether verbs or nouns are more common in English by comparing a large sample of verbs and nouns. You can also compare across languages. Children learn verbs more slowly than nouns in
. Is that because verbs are more common in Chinese than English?
OK, those are fun experiments, but none of them sound very cutting-edge. If you want to see the Google corpus in action, check out
to answer their questions. Google is probably less-commonly used in more formal contexts, but the PsychInfo database turned up 76 hits for "Google." Many were studies about how people use Google, but some were specifically using the Google corpus, such as "Building a customised Google-based collocation collector to enhance language learning," by Shesen Guo and Ganzhou Zhang. Another -- "Nine psychologists: mapping the collective mind with Google" by Jack Arnold -- looked at the organization of conceptual knowledge. At a recent conference, I saw a presentation by vision scientists using Google Image to explore the organization of visual memory. I expect to see more and more of this type of research in the near future.
Of course, there's nothing specific to Google about this. It's just what everybody seems to use.
Enter one address or multiple addresses separated with commas.
I manage the linguistics department of a software company; we use corpora and webskimming in a multitude of languages on a daily basis. We also use google for a lot of minor research tasks. For a while we played with using web counts from both google and yahoo and found that they were just not reliable enough to provide useful data. Turns out there are all kinds of finicking things that the engines do that mess up the results (not the least of which is throwing away all the diacritics!).
has quite extensive commentary on the search engines and their counts which makes for some very interesting reading. Here are a couple of the posts:
http://aixtal.blogspot.com/2005/02/web-googles-missing-pages-mystery.htm...
http://aixtal.blogspot.com/2005/08/yahoo-19-billion-pages.html
Myself, I'd be skeptical of any conclusions drawn from using "google as corpus" unless the questions being asked are not susceptible to the idiosyncrasies of the engine.
what you get is more than just not an island in the pacific or a species of coffee bean, what you get is good solid evidence that the billions of pages stuffed into google are tainted data representing a very narrow demographic. How could any meaningful language research be based on the usage by such a tiny cross-section? A billion pages all written by milk truck drivers from Minneapolis is hardly an indication applicable to the broader population.
And then there's the many pages written by bots. Try the Google "blog" function for any common term, most of the pages you hit are troll sites for casino ads (or worse). Many other terms encounter the word-salad pages trolling for keywords, also to lure clickthroughs into paid page-impressions. It would take a mighty fine bit of AI coding to create a word-sifting algorithm that could account for this noise!
In the old days, people had questions and would say "I guess some day we'll ask God when we get to heaven." I joke now that the wait is over-- questions? Just ask Google. Google has an opinion on anything that is a hot topic or debatable.
Some very good points were raised above. Search engine corpora are not a perfect source. There are some over-represented words. Also, you can't tease apart different meanings of some words (like the "java" example above).
But none of our other sources are perfect either. The gold-standard Brown Corpus was carefully compiled from a set of "representative" sources, but the truth is that in many projects I've done I've found the Brown Corpus too unreliable to use.
So I definitely wouldn't read my post as saying "Google solves all problems." It doesn't. But if we stopped doing science until we had the perfect materials, we'd still be waiting to do Newton's experiments. You work with the best you have, and for some questions, Google is the best available option.
You certainly can ask whether, even in those cases, Google gives you sufficiently reliable information as to be useful. I think it does, but I'm willing to hear arguments otherwise. I look forward to reading the articles linked to above.
OK, I just read the first of those Veronis articles about the Google index. It was some interesting research, but I think he came to the wrong conclusions.
For those who haven't read it, he is puzzled by the fact that sometimes if you search for A or B you get fewer counts than you do if you search for just A. He develops the hypothesis that this is because Google doesn't index as many pages as it says it does.
He then tests this hypothesis in an interesting way. He picks a list of English words and then searches for them in all the Web or just in English. He finds that the "just English" search finds only about half as many pages.
I didn't actually read the end of the post, because at this point I got distracted trying to repeat his results. You can repeat them, to an extent. If you look for "accumulated" in English only, you find half as many pages as on the entire Web. What happens if you look for "accumulated" in Japanese pages? You find huge number of examples! (You also find a decent number on Russian pages -- the only other language I tried.)
What's going on? Do those Japanese pages really have the word "accumulated" on them? Yes, they do.
. I haven't tried every language or done the math, but I am willing to bet that those "missing" pages with "accumulated" will be accounted for in other languages.
Is this strange? Not really. My experience in a range of other countries (Russia, Spain, Taiwan, China, Mexico, Japan) is that a great deal of English has been incorporated into the texts of other languages. While it is unusual to read an article in English and come across a long quote in a foreign language, that's not so uncommon in some other countries. (If you want literary examples, long swaths of War and Peace were written in French.)