We want that spicy info
we’re going to be building a tell analyzer into the site with the game automation stuff so you’ll all be able to do your own research.
meta on the streets, asspull in the sheets
nobody needs to know - just call it “meta”.
Meta read, sheep this
im more interested in whatever it is that @Cactus just ran
yeah me too lol
In my experience sentiment analysis doesn’t test that well for deception.
how does sentiment analysis work in mafia?
if you find some correlation between alignment and type response score then it could be useful.
Potentially strong for certain individuals (who tend to show a certain sentiment analysis as one alignment). Don’t think it’s super strong in general.
I might turn out to be wrong though! iirc someone tested it on a large dataset and found timestamp stuff was better in general.
tbh it might be better if you could have it individualized to each person which ig is what you were trying to do.
That’s bounded by sample size and it’s theoretically a lot more interesting if you can find something fairly strong for all people and then subsets of all experience people/or all new people etc. But yes, if you take all of someone’s posts and their alignments and grind them through something you get interesting results. We’ll be building that into the site. In theory it’ll tell scum what their own tells are.
that has the potential to make our site scum game literally the best on the internet which is super hype.
it’ll be an experiment. a lot might change. kinda like when chess engines dropped.
Agreed! Though of course it depends how ‘deception’ is being defined and whether the analysis is individuated (e.g. detecting a change to passive voice if that’s unusual for that particular person, etc).
Fortunately, sentiment analysis in mafia can be used for a lot more than just predicting alignment.
Interesting! What will you be using?
Probably just text (words and length), post times, and relevant vote info (time between, position etc.)
Yeah I agree this probably works for certain people. Didn’t test that heavily.
leaving as reference: https://nlp.stanford.edu/pubs/sidaw12_simple_sentiment.pdf