NLP: Building (small) corpora, or "Where to get lots of not-too-specialized English-language text files?"

Question

Does anyone have a suggestion for where to find archives or collections of everyday English text for use in a small corpus? I have been using Gutenberg Project books for a working prototype, and would like to incorporate more contemporary language. A recent answer here pointed indirectly to a great archive of usenet movie reviews, which hadn't occurred to me, and is very good. For this particular program technical usenet archives or programming mailing lists would tilt the results and be hard to analyze, but any kind of general blog text, or chat transcripts, or anything that may have been useful to others, would be very helpful. Also, a partial or downloadable research corpus that isn't too marked-up, or some heuristic for finding an appropriate subset of wikipedia articles, or any other idea, is very appreciated.

(BTW, I am being a good citizen w/r/t downloading, using a deliberately slow script that is not demanding on servers hosting such material, in case you perceive a moral hazard in pointing me to something enormous.)

UPDATE: User S0rin points out that wikipedia requests no crawling and provides this export tool instead. Project Gutenberg has a policy specified here, bottom line, try not to crawl, but if you need to: "Configure your robot to wait at least 2 seconds between requests."

UPDATE 2 The wikpedia dumps are the way to go, thanks to the answerers who pointed them out. I ended up using the English version from here: http://download.wikimedia.org/enwiki/20090306/ , and a Spanish dump about half the size. They are some work to clean up, but well worth it, and they contain a lot of useful data in the links.

---

Answer 1

You've covered the obvious ones. The only other areas that I can think of too supplement:

1) News articles / blogs.

2) Magazines are posting a lot of free material online, and you can get a good cross section of topics.

Answer 2

Wikipedia sounds like the way to go. There is an experimental Wikipedia API that might be of use, but I have no clue how it works. So far I've only scraped Wikipedia with custom spiders or even wget.

Then you could search for pages that offer their full article text in RSS feeds. RSS, because no HTML tags get in your way.

Scraping mailing lists and/or the Usenet has several disatvantages: you'll be getting AOLbonics and Techspeak, and that will tilt your corpus badly.

The classical corpora are the Penn Treebank and the British National Corpus, but they are paid for. You can read the Corpora list archives, or even ask them about it. Perhaps you will find useful data using the Web as Corpus tools.

I actually have a small project in construction, that allows linguistic processing on arbitrary web pages. It should be ready for use within the next few weeks, but it's so far not really meant to be a scraper. But I could write a module for it, I guess, the functionality is already there.

Answer 3

• Use the Wikipedia dumps
  • needs lots of cleanup
• See if anything in nltk-data helps you
  • the corpora are usually quite small
• the Wacky people have some free corpora
  • tagged
  • you can spider your own corpus using their toolkit
• Europarl is free and the basis of pretty much every academic MT system
  • spoken language, translated
• The Reuters Corpora are free of charge, but only available on CD

You can always get your own, but be warned: HTML pages often need heavy cleanup, so restrict yourself to RSS feeds.

If you do this commercially, the LDC might be a viable alternative.

Answer 4

Looking into the wikipedia data I noticed that they had done some analysis on bodies of tv and movie scripts. I thought that might interesting text but not readily accessible -- it turns out it is everywhere, and it is structured and predictable enough that it should be possible clean it up. This site, helpfully titled "A bunch of movie scripts and screenplays in one location on the 'net", would probably be useful to anyone who stumbles on this thread with a similar question.

Answer 5

If you're willing to pay money, you should check out the data available at the Linguistic Data Consortium, such as the Penn Treebank.

Answer 6

Wikipedia seems to be the best way. Yes you'd have to parse the output. But thanks to wikipedia's categories you could easily get different types of articles and words. e.g. by parsing all the science categories you could get lots of science words. Details about places would be skewed towards geographic names, etc.