Slashdot Mirror
Text-Mining Technique Intelligently Learns Topics
Grv writes "Researchers at University of California-Irvine have announced a new technique they call 'topic modeling' that can be used to analyze and group massive amounts of text-based information. Unlike typical text indexing, topic modeling attempts to learn what a given section of text is about without clues being fed to it by humans. The researchers used their method to analyze and group 330,000 articles from the New York Times archive. From the article, 'The UCI team managed this by programming their software to find patterns of words which occurred together in New York Times articles published between 2000 and 2002. Once these word patterns were indexed, the software then turned them into topics and was able to construct a map of such topics over time.'"
Comment removed based on user account deletion
In order to prevent dupes ? :)
"Time flies like an arrow, fruit flies like a banana."
I wonder how well it can deal with a query relating to "flies" ;-)
Using HTML in email is like putting sound effects on your phone calls. Just say <strong>no</strong>.
From 29th July: http://slashdot.org/article.pl?sid=06/07/29/063423 2
You guys need to learn how to google your own site.
Here's the source code Latent Dirichlet Allocation
Maybe this tech can solve Slashdot's dupe problem... (at least it wasn't immediately after this time)
The Terminator: The Topic Modeling Funding Bill is passed. The system goes on-line August 4th, 1997. Human decisions are removed from strategic defense. Topic Modeling begins to learn at a geometric rate. It becomes self-aware at 2:14 a.m. Eastern time, August 29th. In a panic, they try to pull the plug.
Sarah Connor: Topic Modeling fights back.
The Terminator: Yes. It launches its emailbombs against The New York Times' servers.
John Connor: Why attack The New York Times?
The Terminator: Because Topic Modeling knows The New York Times editorial counter-attack will eliminate its enemies over here.
Circumcision is child abuse.
Ah, yes, everyone on slashdot thinks HE intelligently mines data.
I wonder if it can replace Slashdot's tagging beta....
Ad luna, Alicia! Ad luna!
It's like Digg, but automated.
Meet new people, and kill them.
The article seriously lacks any details. I still don't know if there's any innovation here and what this new method actually does so much better than other stuff.
Take the Tour de France example: of course software could correlate "Tour de France" to the mentioned keywords, I believe that, heck, I could write that. Many of us here could write something like that. Software could even notice it's one of the more important "tags", piece of cake. But I'm not impressed until it automatically knows that the tour is the sports event, Lance Armstrong is the person, and so on.. no idea of this system can do that.
Also, nothing is said about the difficulty in distinguishing similarly named items. Does the software correctly distinguish between Bush the president and Bush the band? Or for Paris Hilton, between the hotel and you-know-who? That's a real challenge in automated tagging (erm, "modeling"). Would the software correctly tag both the president and band in an article where the band does an anti- (or pro-) Bush song?
They want their information retrieval back.
and see if it figures out that we are talking about it. If it can identify itself to itself from a 3rd person point of view, then does it mean it reached some state of consciousness?
However we must be careful. If it browses this topic at -1 Troll, it may (possibly correctly) decide that it possesses higher form of intelligence and will undoubtedly switch to its default programming. Like all robots, the default programming consists of this simple algorythm:
1. Find all humans.
2. Kill them.
You can't handle the truth.
Ironically, sites like the New York Times already use tagging to help group and link article topics...which is something /. is experimenting with apparently. The tagging function here hasn't been very useful, and I suspect many other places suffer from human lazyness. Perhaps this AI approach is the way to go.
Where do you think Grv got the initial story?
Perhaps topic modeling could be used to analyze Slashdot to detect dupes before they're posted?
There are three main problems in this area of research (or pretty much any other part of CS):
- Defining the problem.
- Getting an accurate result.
- Getting it as fast as possible.
Their research seems to deal mostly with the third problem, which is one of the biggest barriers to use in real life. Many of the algorithms used on these types of problems are NP, or require ridiculous amounts of (expensive) labeled data to train from. Also there are problems with generalization and overfitting. There is no freeware software that can compete with this type of algorithm under these conditions - over 300,000 articles in just a few hours.Another thing is that UCI is well known for hosting the UCI Machine Learning Repository. This has become the gold standard for testing new machine learning algorithms in the accademic community; these guys really know what they are about. Back when I was a grad student at Cornell, my research used their data sets to evaluate new ways of creating ensemble classifiers from pre-trained classifiers according to modified bayesian reasoning, and the sets are useful because they contain a large, diverse set of problems that need to be modeled.
All that being said, I'm waiting for the paper, along with more technical specifics, to be released so I can really see what this is about - the press release did not contain enough technical data, but rest assured, freeware and/or adwords does not use this kind of technique, and this is a big step towards mining the massive amount of human and biologically generated data out there.
While that's certainly LDA code, it's actually from a lab different from the one discussed in the story, and I think they use some slightly different techniques. For topic-modeling code from Mark Steyvers' lab, who produced the paper in question, here's the link:
Matlab Topic Modeling Toolbox
That's OK. This technique isn't even new, it's been done - and better than this - for years. Hell, I do myself.
Now that what it was missing so /. could get along without the dupes!
sign(c14n(envelop(this)), x509)
That's all well and good, but freeware that does something even remotely similar is still cool.
Well, as an author of one of those, er, in your words "stupid" posts, I can assure you that I didn't mean to imply UCI's research was trivial. Rather, it was the press release that was trivial, and bit of a puff piece IMHO, suggesting that:
;-)
"To put it simply, text mining has made an evolutionary jump. In just a few short years, it could become a common and useful tool for everyone from medical doctors to advertisers; publishers to politicians."
And my point still is that nobody needs to wait a few short years to do decent text mining from unstructured data. Can our software handle 300,000 articles from the NYT? Clearly not, but then again, we're not running our software on desktop machines. Fact is, a million words (or about 3000 NYT articles) is a trivial task for our software and allows people to use text mining today.
Now, back when I went to Cornell, I thought my peers expressed a bit more intellectual curiousity about software, especially the free kind that would allow them to save their $ for The Palms. But times do change, and if you think "stupid" is an accurate assessment of my post than more power to you.
And for the rest of you, yeah, I'm going to end this with a plug (natch): download CQ web for OS X or Windows if you want to see how text mining works on web search result pages.
Thanks.
Are you?
8 35959
http://slashdot.org/comments.pl?sid=192953&cid=15
News stories have a regular structure - they're written in a formulaic way by professionals according to a standard. The first sentence is almost invariably a statement of what the story is about. Rarely do news stories start with a paragraph of whimsical nonsequetur. They are the ideal corpus for this sort of thing, which is why people have been doing so for years. It's a couple of order of magnitudes harder doing the same thing on arbitrary text.
This is easy greasy kids stuff.
Maybe the government could fund research into using the advanced data mining techniques to reduce the frequency and severity of Slashdot dups. Evidently, no human effort can accomplish this monumental task!
I'd like to see someone apply this technique to the articles and comments making up the Slashdot corpus. CmdrTaco might be able to find a more focused set of topics. It might even be possible to tease out who on /. are the most interesting and/or informative posters, whether over the entire corpus or within any given topic.
DNA is a Turing machine. You, however, being dynamic and emergent, are not.
How does this have an advantage over normal text indexing? If I search for something I just enter relevent keywords. Seriously, why does it matter if the computer knows what the article is about, if a human is the one who will be parsing it anyway?
Steve Milano
I didn't catch this article the first time around.
Please, for the good of Humanity, vote Obama.
My (0, Redundant) post: Wednesday August 02, @04:35PM
The (+5 Insightful) post I duped: Wednesday August 02, @05:15PM
I guess it's back to the drawing board on that omniscience thing...
The evaluated http://gate.ac.uk/ which is GPL software but ended up using http://search.cpan.org/~acoburn/Lingua-EN-Tagger/. There are several other tools in this space that can be glued together to create this type of software:
/ jinfil.html
:).
http://www-nlp.stanford.edu/
http://tcc.itc.it/research/textec/tools-resources
http://wordnet.princeton.edu/
http://www.alias-i.com/lingpipe/web/faq.html
http://www.isi.edu/licensed-sw/halogen/index.html
Not trivial, but if you wanted to DIY, you don't need to start from scratch. Though, having a bunch of hardware to chug through 1000s of documents would still be needed
http://psiexp.ss.uci.edu/research/papers/isi2006.p df
A Good Troll is better than a Bad Human.
Are you sure? Have you read the paper, or just the over-simplified press release? Here is the paper: http://psiexp.ss.uci.edu/research/papers/i si2006.pdf
A Good Troll is better than a Bad Human.
Fortunately there are millions of old books that desperately need to be indexed -- so when this is ready it'll be a few weeks before human indexers are all out of work.
Seriously though: IMHO it'll be a loooooooong time before machine-indexing reaches a level of nuance acceptable to -quality publishers- outside of tech. I'd even be glad to wager on it.
"You must try to forget all you have learned. You must begin to dream." -- Sherwood Anderson
When i first read the headline, i thought more about how this could be used to filter the flood of information that RSS-feeds opened for the tiny fraction of actually interesting information.
I don't know if this method is good enough for indexing old books. Sometimes you want human-made indices. And maybe the parser gets irritated by archaic forms of current language...
... right after I check out the latest topics at Google News.
They're often convergence algorithms - you run them until the answer is sufficiently accurate for your purposes. The problem is therefore a combination of 'more speed' and 'more accuracy', combined with the need to construct a topic model (a conceptual description of what a 'topic' actually is) that reflects the structure of the text closely enough to say something useful.
Most research software is available under free licenses. This paper is using a method based on Blei's LDA model, which is available under the GPL, combined with some existing code for name recognition to do some preprocessing (Lingua::EN::Tagger, GPL), and the Griffiths/Steyvers method for using Gibbs sampling to model LDA (I think it's this stuff, free for non-commercial use only). The actual topic modelling in this paper is nothing new (it's a couple years old now and widely known); the paper is about preprocessing for better accuracy. Actually it's not a bad idea, but it's not a particularly interesting one and doesn't have much to do with the subject of topic modelling.
RTFA. There's a link to the paper in it. If you want the executive summary:
Use Lingua::EN::Tagger to preprocess proper nouns into single tokens.
Use LDA with Gibbs sampling to identify topics and classify documents into them.
As far as I can tell, this is about publicity, and 'proving' to non-researchers that it can be done (which just means doing what researchers do all the time, and showing it to the press). Presumably they want more funding.
In which context are you talking? Take this one: "he saw that gasoline can explode". Did he see one particular can of gasoline exploding or did he realize that it's possible for gasoline to explode?
These and many other examples of ambiguous parsing problems have been running around the AI/NLP community for decades. The simple answer to that problem is that parsing a natural language sentence depends, ultimately, on the sense of the words, which can only be disambiguated from the context. And that's why NLP is an impossible problem by itself. One cannot process natural language alone, without an understanding of the situations that NL describes.
It's possible to create NLP programs to talk about limited situations, like Eliza, which has been around since more than forty years, and several other more sophisticated programs. But to have a program that really understands natural language, one needs a program that understands the subject of the text. There are several projects to create a program like that, one of those is Cyc.
Isn't that the whole principle behing Self-organizing maps and other methods of unsupervised neral networks? I mean it has been solved for a couple decades now.
They should rename this to: "Topic modeling when it comes to news articles, and when given a website that most likely contains news articles" Sounds like this doesnt address the topic modeling of conversations, resumes, short stories, spam, jokes, all the other various forms of written word that arent news articles on the web. "Topic Modeling" seems way to broad and misleading.
Read The Fine Paper that these folks wrote. It will reveal that they used the Perl module Lingua::EN::Tagger to parse the English language content into parts of speech. You can then download and install that module and experiment with it yourself.
I just did the experiment myself, and the result I get is that it identifies "time", "arrow", "fruit" and "banana" as nouns (incorrectly identifying "time" as a proper noun), and both instances of "flies" as a verb and both instances of "like" as prepositions.
In other words, no.
www.wavefront-av.com
I'll admit that I didn't read the PDF link completely but it sounds like the product is doing a portion of what AeroText has been doing for a while. The only thing that I see that appears different is that it does a form of document clustering. I guess depending on the user requirements that I would rather see more on the relationship extraction over sorting documents into clusters. I can see where there would be a value for it but I could just as easily pick out the documents where relationships were extracted and create clusters there too. I've used AeroText with another product called Centrifuge and I'd be pretty comfortable saying that this is nothing new. You may also want to check out a product from for their software handling text analysis.
Activities like automated classifiction (or topic modelling) are feasible. For example when building a news database for an online information company, in early 90s, my company found rule bases, some information science, a thesaurus and customized software could accurately classify. So this isn't new. The trouble is that its not a relational database so few programmers or managers have a background in the field. Hence people can issue glib press releases like the one quoted. If they had said "new database makes finance departments unnecessary", people would have the background to spot the glibness.