Thank you for sharing your latest efforts and results on computer assisted comment curation. Well knowing the hot mess that some comments can turn into, I hope to see this system become progressively better at handling comments and hopefully their comment removal faster when needed.

The What's next? section shows that you're interested in continuing research with a special focus on sentiment analysis. The way I see it, such a system in practice could have also contributed to the quick detection and clean up of comments which are not constructive. Alas, the comment classification initiative under the Welcoming Wagon only considered a categorization based on "friendliness" (OK, unfriendly, abusive), rather than overall usefulness. In other words, there would have been value in introducing a category for comments which could be removed at any time without losing information, like in the Higgs application at SOBotics. Food for thought nevertheless.

I also have one topic that I would like to see clarified: Are moderators making substantial false positives in their unfriendly comment handling?

This statement caught my attention:

We know that not all of the comments the robot flags are unfriendly and moderators do accept some of those flags because the comment is worthy of removal for other reasons.

This seems to contradict how comment flags from humans are usually handled: they can be declined if the comment should still be removed but is not "hot" enough to merit an unfriendly flag. Some moderators are even known to have a user-script to facilitate this 2-step operation. Do you think that the number of accepted unfriendly flags on non-unfriendly comments in the data set is non-negligible still?