Have you considered that your political bubble might minimize reporting of hate and violence originating from your in-group, while amplifying reporting of hate and violence originating from your out-group?
Yes, it’s definitely a filter I live in. And we unfortunately do not have their public training data corpus.
However, it’s also true that I can trivially find clips of mainstream republican leaders over the last few years with violent charged rhetoric, and it is comparatively harder to find such rhetoric from mainstream democrats[0][1]
For many of the other groups, hate has been historically so widespread and common there are whole organizations like the annti defamation league organized around combating it. There is no way that even if you believe today people are furiously writing hate against Christians into twitter 24/7 or wherever my bubble they could have caught up with the historic use of hate speech against Muslims in the wake of 9/11. And so on for the other disadvantaged groups.
And all this is not to say it isn’t a potential problem or a potentially useful metric. It could be that synthetic hate is needed to anticipate new kinds of unlikely and rare sentence structures that might arise. If this is true it also reminds us that these models will need constant training as new kinds of hate speech become more popular and that it will always be lagging.
