There’s a lot of scholarly work on HTML content extraction, e.g., Gupta & Kaiser (2005) Extracting Content from Accessible Web Pages, and some signs of interest here, e.g., one, two, and three, but I’m not really clear about how well the practice of the latter reflects the ideas of the former. What is the best practice?
Pointers to good (in particular, open source) implementations and good scholarly surveys of implementations would be the kind of thing I’m looking for.
Postscript the first: To be precise, the kind of survey I’m after would be a paper (published, unpublished, whatever) that discusses both criteria from the scholarly literature, and a number of existing implementations, and analyses how unsuccessful the implementations are from the viewpoint of the criteria. And, really, a post to a mailing list would work for me too.
Postscript the second To be clear, after Peter Rowell’s answer, which I have accepted, we can see that this question leads to two subquestions: (i) the solved problem of cleaning up non-conformant HTML, for which Beautiful Soup is the most recommended solution, and (ii) the unsolved problem or separating cruft (mostly site-added boilerplate and promotional material) from meat (the contentthat the kind of people who think the page might be interesting in fact find relevant. To address the state of the art, new answers need to address the cruft-from-meat peoblem explicitly.
Extraction can mean different things to different people. It’s one thing to be able to deal with all of the mangled HTML out there, and Beautiful Soup is a clear winner in this department. But BS won’t tell you what is cruft and what is meat.
Things look different (and ugly) when considering content extraction from the point of view of a computational linguist. When analyzing a page I’m interested only in the specific content of the page, minus all of the navigation/advertising/etc. cruft. And you can’t begin to do the interesting stuff — co-occurence analysis, phrase discovery, weighted attribute vector generation, etc. — until you have gotten rid of the cruft.
The first paper referenced by the OP indicates that this was what they were trying to achieve — analyze a site, determine the overall structure, then subtract that out and Voila! you have just the meat — but they found it was harder than they thought. They were approaching the problem from an improved accessibility angle, whereas I was an early search egine guy, but we both came to the same conclusion:
Separating cruft from meat is hard. And (to read between the lines of your question) even once the cruft is removed, without carefully applied semantic markup it is extremely difficult to determine ‘author intent’ of the article. Getting the meat out of a site like citeseer (cleanly & predictably laid out with a very high Signal-to-Noise Ratio) is 2 or 3 orders of magnitude easier than dealing with random web content.
BTW, if you’re dealing with longer documents you might be particularly interested in work done by Marti Hearst (now a prof at UC Berkely). Her PhD thesis and other papers on doing subtopic discovery in large documents gave me a lot of insight into doing something similar in smaller documents (which, surprisingly, can be more difficult to deal with). But you can only do this after you get rid of the cruft.
For the few who might be interested, here’s some backstory (probably Off Topic, but I’m in that kind of mood tonight):
In the 80’s and 90’s our customers were mostly government agencies whose eyes were bigger than their budgets and whose dreams made Disneyland look drab. They were collecting everything they could get their hands on and then went looking for a silver bullet technology that would somehow ( giant hand wave ) extract the ‘meaning’ of the document. Right. They found us because we were this weird little company doing “content similarity searching” in 1986. We gave them a couple of demos (real, not faked) which freaked them out.
One of the things we already knew (and it took a long time for them to believe us) was that every collection is different and needs it’s own special scanner to deal with those differences. For example, if all you’re doing is munching straight newspaper stories, life is pretty easy. The headline mostly tells you something interesting, and the story is written in pyramid style – the first paragraph or two has the meat of who/what/where/when, and then following paras expand on that. Like I said, this is the easy stuff.
How about magazine articles? Oh God, don’t get me started! The titles are almost always meaningless and the structure varies from one mag to the next, and even from one section of a mag to the next. Pick up a copy of Wired and a copy of Atlantic Monthly. Look at a major article and try to figure out a meaningful 1 paragraph summary of what the article is about. Now try to describe how a program would accomplish the same thing. Does the same set of rules apply across all articles? Even articles from the same magazine? No, they don’t.
Sorry to sound like a curmudgeon on this, but this problem is genuinely hard.
Strangely enough, a big reason for google being as successful as it is (from a search engine perspective) is that they place a lot of weight on the words in and surrounding a link from another site. That link-text represents a sort of mini-summary done by a human of the site/page it’s linking to, exactly what you want when you are searching. And it works across nearly all genre/layout styles of information. It’s a positively brilliant insight and I wish I had had it myself. But it wouldn’t have done my customers any good because there were no links from last night’s Moscow TV listings to some random teletype message they had captured, or to some badly OCR’d version of an Egyptian newspaper.
/mini-rant-and-trip-down-memory-lane