Building a conversation corpus by text derivation from “germ dialogs”