<p>In the field of knowledge building, “Discourse” plays a central role. If real-time and precise classification, as well as effective monitoring and intervention, can be achieved for the various discourses generated during the process of online collaborative knowledge building, it will greatly promote “idea improvement”. However, knowledge building systems rely heavily on the semantic recognition and judgment of natural language. Thus, existing knowledge building systems augment the context through external resources such as knowledge graphs and model entities based on their interrelationships. Nevertheless, these methods overlook the rich intrinsic information within entities. To address this issue, we introduce a knowledge graph-based enhanced entity representation learning framework. This framework leverages knowledge graphs and pre-trained language models to improve the semantic understanding of entities. It effectively captures the temporal information of entities in conversations using positional encoding. Through this framework, the accuracy of discourse classification in the knowledge building process is significantly enhanced. Comparative and ablation experiments demonstrate that the SVM algorithm outperforms other methods, achieving an accuracy of 92% and an F1 score of 87% in discourse classification. This method holds significant value in the context of online collaborative knowledge building, pedagogy and technology. It effectively supports the development of process evaluation and precisely regulates learners’ knowledge building processes. Furthermore, it promotes the continuous enhancement of discourse, providing robust support for the high-quality development of online collaborative knowledge building instruction.</p>