Bulletin boards require moderation. Even on serious bulletin boards for professionnals, moderation is mandatory. You need to remove spam, defames, abuses, threats, hateful statements, discussions about illegal activities and vulgar, obscene, pornographic, or indecent language.
Spam filters are text classifiers. They can tag spam and ham. Could spam filters also be used to moderate bulletin boards?
This is something I wanted to try for a long time. Some previous experiments with DSPAM were interesting, but there were many false positives and I didn’t try any further.
But I recently seriously implemented that idea. The bulletin board is based upon Vbulletin and the text classifier is CRM114.
A daemon is scanning new messages, with an intentional delay, in order to avoid people who would post the same message over and over again as it disappears. These messages are processed by CRM114. If they are detected as spam, they become invisible, actually queued for manual moderation. A manual moderation interface shows unmoderated messages. The moderator can let a message pass through. In that case, and if the message was previously unclassified, CRM learns the message as non-spam. If the moderator deletes a message that was either a false negative or an unclassified message, CRM learns the message as spam. Quite basic, although the actual details are a bit more complex in order to deal with meta-moderation and with users who intentionnally remove some messages.
So, how does this system perform? Amazingly well, I was really impressed. It works and CRM114 learns very fast.
Illegal message are immediately trapped by CRM114, this is truely amazing. I was shocked to discover that it was quite effective to find aggressive messages. I was also shocked by the tiny number of false positives. At that point, and for that specific task, the classifier seems to be quite useful.
Of course, this doesn’t replace human moderators. Because moderators have to move threads into the right categories, to help users, to understand the real meaning and the potential implications of some posts, etc. But the system catches most illegal messages before the manual moderation. It’s a very efficient proactive tool, especially when you can’t afford a team of moderators working 24 hours a day, 7 days a week.
I was also simply amazed by CRM114 itself, although I was a die-hard DSPAM lover. CRM114 is way, way, way more flexible, it’s fast, it’s actively maintained and the css files are reliable.