The paper is here sarcasmAmazonICWSM10.
The team at Hebrew University first scanned through 66,000 product reviews at Amazon and developed a classifier to group patterns of text into different sarcastic classes. They then seeded their machine-learning algorithm with subset of these samples. Apparently, the algorithm is in on the joke 77% of the time.
This is similar to what I’m hoping to do for disagreements. Amazon reviews, contentious blog posts and, even mailing lists for technology standards, provide a rich textual corpus to play with.