The ideal, platonic version of this would be that moderators only see an "identity scrambled" version of each tweet/post when they make their moderation decision. Like a screen that blinds orchestra musicians when they audition, the human would see a statement like "I hate New Yorkers" and not know if the original message said "I hate New Yorkers" or "I hate Floridians." So they would have to make a decision based on the general principle of whether a statement of this form is allowable.
Anywhere you want to draw the line is fine with me, as long as you draw it consistently.