Continuing the discussion from Trying an AI/ML sentiment analysis system:
Discourse has replaced their first experiment in AI sentiment analysis with a more powerful suite of tools. All of these are using open source models and running entirely in their infrastructure (the same infrastructure that runs this site).[1]
You can read details here: Discourse AI - plugin - Discourse Meta, but the summary of what I’ve turned on is:
-
Toxicity module: automatically flags posts classified as “rude, disrespectful, or unreasonable”, with adjustable thresholds for: toxicity overall, insults, threats, identity attacks, and sexually explicit comments. The plugin can chide users or even hide posts, but I’ve set it to just flag. That means everything will get a human review.
I’m a little skeptical of how well this will work, to be honest, and intend to set the threshholds so that false positives are rare. But I hope to be pleasantly surprised.
-
Sentiment analysis: This one just provides reports on posts classified across “positive / negative” and “emotion” (joy, surprise, anger, disgust, fear, sadness, or neutral).
I’m also a little skeptical about this, especially for technical posts. But again, let’s see what it looks like!
-
NSFW Image Detection: Currently, new users are limited in the number of images they can post in order to prevent problems here.[2] If this works well, we could significantly relax that so people posting screenshots and design ideas and so on aren’t frustratingly hampered.
-
Semantic Related Topics: Previously, the “Related Topics” at the bottom of a post were actually pretty much random and not likely very related. This might make those actually useful.
-
Semantic Search: Same thing for search. This doesn’t replace the existing esarch, but adds a new Topics (AI Search) option to the search page. I’m very interested in feedback on how this works.
-
Summarize topics: Adds a “Summarize via AI” option to long threads. Useful, or a gimmick? Probably better than the current approach of picking posts by the number of
s they get. _This one does not seem to be working yet — maybe takes a little while to index some things? Or maybe I’m just missing something in the configuration.