Enhancing Language Models in Statistical Machine Translation with Backward N-grams and Mutual Information Triggers