Mental health education in colleges and universities has made considerable progress, but the existing assessment model still faces challenges in terms of time overhead and rank indicators. In response, this paper proposes a new psychological education assessment model for colleges and universities, based on multimedia feature extraction techniques. The proposed model utilizes word vectorization with Word2vec and an improved transformer network, incorporating deeply separable convolutions and an LSTM network to establish long-range dependency coding. Experimental results show that the proposed model outperforms traditional feature extraction methods in terms of extraction speed, feature readability, and model efficiency. Furthermore, the study suggests that the use of reinforcement learning can improve the system's ability to capture key concepts and enhance accuracy. The proposed approach has significant implications for improving mental health education in colleges and universities and can be applied to similar professional environments.