Efficient language model look-ahead probabilities generation using lower order LM look-ahead information