Abstract: The evaluation of teacher's teaching quality is a very important link of the educational administration in colleges and universities, which aims chiefly to examine the teachers and their teaching activities. Through evaluation the teaching administration can learn about the actual situation about the students and spot problems existing in teaching before it is too late and the teachers can also reflect upon their teaching process so as to improve their teaching quality. Whether a teaching quality evaluation is fair or not can decisively tell whether a teaching evaluation system is workable. A question of questions faced with today's evaluation systems is how to guarantee the fairness. It is proposed in this paper that a neural-network-based online teaching assessment system be designed to delimitate and filter the errors caused by students' capricious grading through neural network training score weight.