Those are good points. But the more important issue will be context. For example, in order for sentiment to be 'racist' - the ethnic nature of the participant

Those are good points. But the more important issue will be context. For example, in order for sentiment to be 'racist' - the ethnic nature of the participants may have to be known. So it's not about 'analyzing a statement' so much as it is understanding so much about people, individuals, current thresholds for tolerance, current events. Etc. etc.. It's impossible to look at a 'post' in total isolation, it has to be put in the context of the real world which is way beyond AI today. So it's not possible to 'train' AI simply by putting 'sentiment' as an input parameter. You'd have to 'train' NN's with inputs that include the entirety of CNN, Wikipedia and pop culture!