I did a project for a class doing stylometry with forum posts. It's harder than blogs because you generally have fewer words per author and some people post disproportionately, but even with a limited feature set, you can achieve much better performance than random chance. Some other people in that class did a similar thing with chat logs and got very impressive results; they used a much larger feature set (word-based multinomial event model, if I recall).
This is a powerful tool; I'd bet that most governments invest heavily in it.
]]>wg
]]>Few know me and my style is varied
I output at least in two languages
In film, in music, in books, on tv
There are a great many instances
of me
And noone has cottened onto yet.
And that's exactly how it should be.
I agree about coding style. I found source code online for the Apple /// OS recently, originally written in 1980-81.
After several minutes of examining the 6502 assembly-language source code of the floppy disk driver, I recognized it as my own. (To be fair, I should admit that I recalled rewriting that driver, though I wasn't sure at first whose version I'd found.)
]]>I lost the artical but Chinese authorities do the same thing with db that store everything political dissidents under nyms write to compare to their university papers so they can match names
]]>Real names aren't required, but please give us something to call you. Conversations among several people called "Anonymous" get too confusing.
]]>Years ago, the comment form of you blog advised > under the field "Name".
Why did you stop this message ? Did a Three Letter Agency tell you that their stylometric software would perform better without this advice ?
Is it a move to fight comment-spammers ?
]]>Writeprint is a real product deployed by Dark Web in 2007, already doing stylometry. Against islamists activists. It was called Writeprint. Quoting http://www.nsf.gov/news/news_summ.jsp?... :
One of the tools developed by Dark Web is a technique called Writeprint,
which automatically extracts thousands of multilingual, structural, and
semantic features to determine who is creating 'anonymous' content online.
[...] By analyzing these certain features, it can determine with more than 95
percent accuracy if the author has produced other content in the past.
@MingoV
Not 5,000 subjects. 5000 words per author [1].
]]>Not soon, it happened a few years ago in Berlin, Germany. A scientist was arrested because the police said, he is a member of a group which burns down army stuff (i think they destroyed just trucks, cars and other vehicles).
One reason they said it, was because some words he was using in his publications, were in the writings of the anti army group.
Here are more informations. His wife made a blog about this:
]]>The identification rate may improve when the pool of writers is reduced, such as when there is a specific group of suspects.
It's always going to be unreliable, perhaps better used as an exculpatory process. If the suspect has a 90 IQ and the text has a high degree of literacy than it's probably not them.
The usual simple method of anonymisation mentioned is to use Google Translate to change your text into different languages, then back again and clean up the obvious mistakes.
]]>And something similar was used on Dexter during the second(?) season when he culled various works for his manifesto and was called out on it because there were dozens of different authors.
]]>Put another way, does stylometry assume authors are not actively trying to defeat it?
]]>