[xXOpkillerXx]

Originally Posted by aperson

I think most societies have a pretty consensus-driven view of what is an insult or not. You could use a population sample to review conversations and determine if there is much of a consensus on whether an insult occurred or not. If this is the case, you can use the responses to train a supervised neural network (see: backpropagation).

The more interesting question becomes what sort of richness do you need in the dataset to make the evaluation? Do we need information on the surrounding scene? Speaker and audience relationship? What happens on different media like video vs text transcripts? Etc...

If there was such a dataset, you'd need contextual information. Otherwise, you'd only pick up insults with very strong words/n-grams such as "fuck you", "bitch", etc. Pretty sure a PCA would show some interesting relationships between the vocabulary and the various contextual data, but then there is just So much info that can be needed to understand an insult...