Kind of playing devils advocate here, but would you not expect your probabilistic hatespeech detector to score "common hatespeech" higher? If 90% of your hates

Kind of playing devils advocate here, but would you not expect your probabilistic hatespeech detector to score "common hatespeech" higher? If 90% of your hatespeech focusses on disabled promiscuous jewish homosexuals, would it not be expected for the hatespeech detector to "perk up" when talking about them? Because the conditional probability that ANY paragraph about jewish homosexuals is hatespeech IS, in fact, increased? IF you wanted your detector to be unbiased, you would have to train it on fictional, unbiased hatespeech (or remove that bias otherwise)-- decreasing its real world performance! edit: That viewpoint would predict that the hatespeech detector "favors" groups that are actually frequently hated on, NOT groups that "californian progressives" deem worthy of protection. That seems plausible to me; atheists, for example, are NOT favored even though they should be much more "californican" than catholics or mormons.