Online one-on-one class is created for highly interactive and immersive learning experience. It demands a large number of qualified online instructors. In this work, we develop six dialogic instructions and help teachers achieve the benefits of one-on-one learning paradigm. Moreover, we utilize neural language models, i.e., long short-term memory (LSTM), to detect above six instructions automatically. Experiments demonstrate that the LSTM approach achieves AUC scores from 0.840 to 0.979 among all six types of instructions on our real-world educational dataset.