Well in the article it stated that "Siri brings in 1 billion queries per week from users" so that's nothing to sneeze at. But the real challenge is the majority of those queries are probably pretty simple queries that Siri doesn't/can't learn from - she already answers them fine. You need alternative data sets (such as the one generated by mechanical turk) to really learn anything new.

But datasets are only 1 part of the problem - innovate, sophisticated and tuned approaches to solving these problems often make the difference.