A sarcasm detector a couple of my buddies and I made for a hackathon in 2018, taking 3rd out of 60 teams.
We used a combination approach, a character RNN + a naive bayesian classifier to determine if a sentence was sarcastic or not.
Honestly, it worked surprisingly well.
We tried it first with just the RNN, and while it worked reasonably well, it had some odd edge cases.
We didn't have enough time to train it again, or generate a better dataset, so I cooked up a simple naive bayesian classifier in Lua, saw that it had comparable success, and averaged the results.
That approach worked excellently.