Improved clustering techniques for class-based statistical language modeling