You will not be able to efficiently automate finding trolls with an algorithm. You may accidentally catch some trolls with it, but this is not really the same as the algorithm <i>working</i> - with a large enough userbase you could randomly ban accounts and probably catch a few trolls as well, given a flexible enough definition of "troll."<p>What you will accomplish with an automated filter is a way for trolls to game the system (that is what actual trolls do), and an excuse for human mods, if you have them, not to care (or worse, to game it themselves in order to censor legitimate posters.) "Future banned user?" How wonderfully fascistic. If you can read that term and feel it's a good way to describe new users to your whatever community, do everyone else a favor and, please, give someone else the banhammer.
Du Plessis said, give me 5 sentences written by the most honest of men and I will find something in them to hang him.<p>The more things change, the more they stay the same...
This is <i>astonishingly</i> stupid. It's also time to retire the word "troll", it has become so bastardized that its use shouldn't be taken at all seriously anymore.
It's not just "trolling". the text describes it as antisocial behaviour (defined as "undesired behavior, which includes
trolling, flaming, bullying, and harassment").<p>To me it seems like detecting "trolling" could be as difficult as detecting sarcasm. It requires information about the context and complicated logic. "flaming", "bullying" and "harassment" usually follow much simpler patterns (language, targeting specific users).
Discussed here:
<a href="https://news.ycombinator.com/item?id=9398399" rel="nofollow">https://news.ycombinator.com/item?id=9398399</a><p>It's not getting much love from the HN community, and has tripped the "flame war" detector - only 26 points, but 40 comments. Much back and forth, much heat, not much light.
In a company situation you also experience trolls and you can't get rid of them, at all. To my personal surprise I find that they, between periods of stress, also contribute to the company. You just need to get around their trolling and they can be a productive part of the community.<p>Maybe instead of detecting them faster we should learn how to integrate them better. If they troll more when resisted (as suggested in the article) then maybe they troll less when supported? They are humans, and as humans they desire to belong, just as we do, right?
Given the stated accuracy and such a low base rate it looks like it would yield ten times more false positives than true positives. To eliminate some of the trolls you effectively eliminate anyone outside the consensus trance. You might as well use astrology.<p>But very interesting to see the actual CNN, ING and Breitbart numbers.
How come so many places don't use a combination of account age and reputation to score and sort posts?<p>For example why hasn't Google tried this with YouTube comments?
The concentration of inside-the-box thinking and bullshit science in this is unbearable, I think it gave me cancer.<p>So was this actually written for April Fools' Day?<p>arXiv:1504.00680v1 [cs.SI] 2 Apr 2015