A statistical approach to language modelling for the ATIS task