With digitized text from five million books, one is never at a loss for words
Googling the Lexicon
The data-driven approach to language studies got a big boost last winter, when a team from Harvard and Google released a collection of digitized words and phrases culled from more than five million books published over the past 600 years. The text came out of the Google Books project, an industrial-scale scanning operation. Since 2004 Google Books has been digitizing the collections of more than 40 large libraries, as well as books supplied directly by publishers. At last report the Google scanning teams had paged their way through more than 15 million volumes. They estimate that another 115 million books remain to be done.
At the Google Books website, pages of scanned volumes are displayed as images, composed of pixels rather than letters and words. But to make the books searchable—which is, after all, Google’s main line of business—it’s necessary to extract the textual content as well. This is done by the process known as optical character recognition, or OCR—a computer’s closest approximation to reading.
In 2007 Jean-Baptiste Michel and Erez Lieberman Aiden of Harvard recognized that the textual corpus derived from the Google Books OCR process might make a useful resource for scholarly research in history, linguistics and cultural studies. There are many other corpora for such purposes, including one based on a Google index of the World Wide Web. But the Google Books database would be special both because of its large size and because of its historical reach. The Web covers only 20 years, but the printed word takes us back to Gutenberg.
Michel and Aiden got in touch with Peter Norvig and Jon Orwant of Google and eventually arranged for access to the data. Because of copyright restrictions, it was not possible to release the full text of books or even substantial excerpts. Instead the text was chopped into “n-grams”—snippets of a few words each. A single word is a 1-gram, a two-word phrase is a 2-gram, and so on. The Harvard-Google database includes 1-, 2-, 3-, 4- and 5-grams. For each year in which an n-gram was observed, the database lists the number of books in which it was found, the number of pages within those books on which it appeared and the total number of recorded occurrences.
The n-gram database is drawn from a subset of the full Google Books corpus, consisting of 5,195,769 books, or roughly 4 percent of all the books ever printed. The selected books were those with the highest OCR quality and the most reliable metadata—the information about the book, including the date of publication.
A further winnowing step excluded any n-gram that did not appear at least 40 times in the selected books. This threshold, cutting off the extreme tail of the n-gram distribution, greatly reduced the bulk of the collection. Combining the 40-occurrence threshold with the 4 percent sampling of books, a rough rule of thumb says that an n-gram must appear in print about 1,000 times if it’s to have a good chance of showing up in the database.
The final data set covers seven languages (Chinese, English, French, German, Hebrew, Russian and Spanish) and counts more than 500 billion occurrences of individual words. The chronological range is from 1520 to 2008 (although Michel and Aiden focus mainly on the interval 1800–2000, where the data are most abundant and consistent).