Strategies for statistical spoken language understanding with small amount of data - an empirical study