Robust Analysis of Spoken Input combining Statistical and Knowledge-based Information Sources