Question about the Automated detection of verbal abuse

FenricsShield·Aug 25, 2016·2 votes·945 views

http://i.imgur.com/VWdW1Qy.png

Generally, I have been a fan of the automated verbal abuse detection system. I understand the way neural networks work, and reporting gives the net a great training set to detect toxic language. But is the net set up such that it can recognize verbal abuse spread out over many lines of text, such as this?

If not, can this be looked into?

If not, can we have tribunal back? :/ This is obvious to a human, but hard for computers to pick up. I imagine there are other ways of obfuscating stuff from the neural net responsible for moderating toxicity reports.

((Note: last time I posted this, it got deleted because I forgot to scrub the names from the screenshot. My apologies; my intent was not to "bring attention to" the person doing this posting, but to raise a discussion on the possible avoidance of triggering the Verbal Abuse detection system.))

6 Comments

ModWuksAug 25, 20164 votes

Post-game chat isn't included in the Instant Feedback system's analysis anyway. For this, you should note that a player was abusive in post-game chat to potentially have the chat logs manually reviewed by the Player Behavior team.

FenricsShieldAug 25, 20162 votes

Didn't know this. When the abuse happens in post-game chat, I'll make sure to mention that in my comments.

Bard R FridayAug 25, 20161 votes

There's no way we can have a computer that completely censors swear words, there's always a way to bypass it. You just need to report them and if they keep doing it they get banned.

FenricsShieldAug 25, 20161 votes

Actually, a neural net is a genius way to do it, because with enough reports, the system can learn all the most frequent ways to misspell or otherwise obfuscate profanity or abuse. I've had enough Machine Learning classes to understand how it works, and it's a genius solution to the problem.

That is, as long as the system tokenizes it properly. If the computer looks at each line individually, then there's not a chance it can find the abuse. After all, "S" is not verbal abuse. "U" is not verbal abuse. "C" is not verbal abuse...