Extending n-gram language models based on equivalent syntactic patterns