If a conversational bot can get to the level of Aspergers (I know, I know, who the hell cares if it's in DSM-5) then I'll be awfully happy.
Though...if you're using it less for conversation, and more for fact scraping, I guess that makes things way more important. Trying to think of how people recognize sarcasm, and it seems to be "a person is stating--without expressing uncertainty or hedging--something that I have good reason to think they don't believe". It seems like every culture also has their own shorthand for sarcastically agreeing with someone, but the shorthand isn't always recognizable cross-culturally (Americans visiting England have a big problem here).
The big AI problem here is, you have to be able to model the writer's thoughts and guess what their opinions are. Ask your stereotypical left-winger and right-winger about gun control and they might both say "Oh, yeah, that's exactly what we need", but you need to be able to predict their actual opinions in order to know if there's dissonance or not. Makes it hard to detect sarcasm in people you don't know anything about, unless they're using shorthand ("yeah right", "great idea--let's discuss it later")