The latter can be learned from "statistical correlations in text", just like Ilya said.