Is AML using a deep learning architecture? I'm sure there are enough phrases that when used alongside or near other phrases can cause trolling to occur. Given a big enough training set and deep enough architecture, I can imagine it would work well. Perhaps you'd want to use a thesaurus to canonicalize words in case your training set isn't big enough. Removing fill words might help as well.

For example, the phrase "Trump was bombastic in his political speech" might become something like "Trump grandiose political speech" or even just "political speech".

I wonder if this sort of feature extraction has been done elsewhere and could be re-used for this problem.

edit: looks good: http://blog.mafr.de/2012/04/15/scikit-learn-feature-extracti...