The generation of meaning in discourse is regarded as the process of knowledge-building, which can be analyzed by semantic waves. Combining the deep learning systems and polynomial fitting method, we present a method to compute multimodal semantic waves. With this method, four representative examples are discussed for the knowledge-building of Chinese, English, Mathematics and Science subjects in Chinese primary schools. From the multimodal semantic waves of these subjects, we can clearly observe the differences between the processes of knowledge-building. Besides, it is advantageous for the proposed method that the value of multimodal semantic waves at any given future time can be predicted by an nth-order polynomial, which may help to study practical classroom discourse.