MTLM: Incorporating Bidirectional Text Information to Enhance Language Model Training in Speech Recognition Systems