Yeah - I looked for the paper that won him the Amdocs prize but couldn't find it. All reports seem to be, um, based on this story. Which is where I found he trained the system using two Fffacebook pages:
posts on Hebrew-language Facebook pages that are almost pure opinion, called “superior and condescending people” and “ordinary and sensible people.” The pages are basically forums for people to let off steam about things and events that get them mad, a substitute for actually confronting the offending person. Between them, the two pages have about 150,000 “likes,” and active traffic full of snarky, sarcastic, and sometimes sincere comments on politics, food, drivers, and much more.
“Now, the system can recognize patterns that are either condescending or caring sentiments and can even send a text message to the user if the system thinks the post may be arrogant,” explained Saig.
System Alert - Possible Arrogance Detected - user message issued
[ 328.0081004] Overtones Warning (bug): Optional FUBAR field Gpe1Block has zero address or length: 0x000000000000102C/0x0 (Sarcasm overflow)
So it's a startup pitch - expect optimistic projections of outcomes. It's even possible (would it detect that) it's based on pure supposition - you know, like maybe the opinion of the machine learning program matched a readers take on those Fffacebook pages.