Computational linguistics
Some notes for now. Note these are from a grad school level background and they're largely aimed at myself in case I ever want to write this up in the main article, while still at least perhaps pointing others vaguely in the direction of stuff to look for if they want to read the literature. But in terms of wikification, I haven't gone for the right tone, for example, and it uses research jargon, and it's too detailed.
The relevant field is called "sentiment analysis" and of course it's occured to them to look at flames too, although the main commercialisation push is towards aggregating opinion. ("85% of people think that Hawaii is a great place to holiday/you should see the new Terminator film".) Partly that's because of the money, partly it's that being wrong about the details of aggregation is a less big deal. (Do you really care if it was only 82% of people who liked Hawaii?)
Bo Pang and Lillian Lee (2008), Opinion Mining and Sentiment Analysis is the standard review of the literature (Bo Pang did a lot of the popularisation of the task as her PhD work).
Sentiment analysis is well known to be a harder task for state of the art text classification techniques. Note that these do go well beyond the individual word-based features used in Bayesian spam classificiation, it's just that that's computationally cheap and happens to work well enough for spam. Given that sentiment analysis is generally one of the harder classification tasks, we can take it at least as a hypothesis that "the same" techniques used for spam aren't going to have as good results.
In terms of a 'harassment' v 'not-harassment' classifcation task, you have the same problems with spam and probably some additional ones:
- false negatives (abuse that gets through the filter) is much higher cost than spam false negatives
- false positivies (legit communication that gets caught in the filter) is at least as high cost as spam false positives, and without doing the experiment you can't be sure you get the same low numbers of false positives, in fact, assume not
Notes on what you'd want to do in terms of the experiment:
- have a large corpus of texts representative of different categories that you find useful ("not abuse" versus "abuse", "not abuse" versus "splaining" versus "abuse"? there's no especial reason to just have two except that because you'll guess right 50% of the time you won't make so many errors near the edges). These should be representative in the sense of being about the right proportions as in the real scenario (which right away suggests you won't get a global solution by training once)
- training, development and test splits
One additional problem you're going to run into here is that for a lot of people (although not all) abuse is relatively rare. I think it's fair to say that most techniques in text classification are designed to distinguish classes of texts that are all reasonably common in the sample.
Thayvian 21:39, October 28, 2011 (UTC)