An integrated semantic analysis system is presented, and the structured language models are proposed based on it. The semantic analysis system can automatically tag semantic class for each word and analyze the semantic dependency structure between words with the precision of 90.85% and 75.84% respectively. In order to describe sentence structure and long-distance dependency, two kinds of structured language models are examined and analyzed. Finally, these two language models are evaluated on the task of Chinese speech recognition. Experiments show that the best semantic structured language model?headword trigram model?achieves 0.8% absolute error reduction and 8% relative error reduction over the trigram model.