ON-LINE ADAPTATION OF SEMANTIC MODELS FOR SPOKEN LANGUAGE UNDERSTANDING