Slashdot Mirror
The New York Times Is Expanding Comments With the Help of Google's AI (recode.net)
An anonymous reader shares a Recode report: The New York Times says it is going to expand the availability of online comments from 10 percent of articles to 80 percent by the end of the year, without adding more moderators to its staff. How are they going to do this? With a machine-learning algorithm, of course. The Times today is rolling out a new structure of comment moderation using software from Google called Perspective, developed by the company's incubator, Jigsaw. The Moderator tool will automatically approve some comments and help moderators wade through others more quickly.
There is a link in the article to here where you can input comments that the system will judge to be 'toxic' or not. There is no sarcasm, irony and bullshit detection that I can tell, only a score that is generated by the combination of keywords used.
For example, "The cake is a lie" receives a 50% toxicity score, "The cake is bullshit" receives %90, and "There is no reason to believe the cake exists." is scored 3%. This system merely weeds out the laziest of trolls.
It's really easy to fool the system to let clearly offensive comments through. It's fooled by simply misspelling words that are deemed offensive, which essentially puts it on the level of Slashdot's l4meness filter (more on this later). Consider the following text, "I don't like [n-words]" that I can't even put in a Slashdot comment without triggering the l4meness filter. With the actual n-word, the Perspective API indicates that it's 87% likely to be perceived as toxic. However, replacing the i in the n-word with ii or a 1 lowers that score all the way to 13%. The same simple tricks that Slashdot trolls use to evade the l4meness filter also work to fool the Perspective API. If the m0deration is automated and people aren't reviewing the comments, it can easily be fooled in its current state.
Here's an experiment that I've been trying. Find any Slashdot article and paste in comments that have been modded up versus comments that have been modded down. Aside from the most blatant of personal attacks, it does a lousy job of identifying which comments are at -1 and have been deemed toxic by the human moderators here. The system can also be defeated by posting something offensive (like the n-word) followed by intentionally benign comments to lower the score. It reminds me of back in the day when trolls would add random text to the end of their posts to defeat Slashdot's l4meness filter.
I'm not convinced that it's any better than the very simple approach of just having a list of banned words. In fairness, it's early in development, but right now the system is a very complex way of implementing an easily defeated barrier against flame wars.
Automated systems are also prone to false positives. My comment is an example. Apparently saying the word "lameness" more than twice triggers the lameness filter. This comment is an example of the problem. I haven't been able to trigger a lot of false positives with the Perspective API, so perhaps that's an advantage.
Ironically, my captcha is "accepted."