As a concept I think this is interesting. Could actually be great as a plugin similar to e.g. Grammarly, where you can automatically do this for all your comment fields.

P.s. As a red-green color blind person I don't see a difference between the two examples (might also be my phone). An additional visual indicator might be helpful.

I hadn't considered that. Will definitely look into some other sort of indicator.

I embedded the sentiment analyzer from Stanford core nlp into a web crawler I wrote, with the idea of instructing the crawler to follow links surrounded by only positive or negative text. Didn't seem all that useful at the time, but now I wonder if it's worth digging up?

I wonder if you could not feed, e.g. reddit comments and identify long chains of respectful conversations for instance.

It could also be a useful moderation tool by pointing out who first started to troll and devolved a conversation into a fece flinging contest.

You may want to train it on some new data then. The data CoreNLP comes trained with is basically yelp review type of stuff, so people talking about a product. Seems to work fine for people talking about other people too. But I think you will get much better results if you train it to actually detect trolls.

Maybe set up a crowdfund to make the internet's first troll content training corpus.

Try to predict karma scores then?