Computer-supported collaborative learning (CSCL) has been an effective pedagogy in the field of education. However, productive collaborative learning often does not occur spontaneously, and learners often have difficulties with collaborative knowledge building and socially shared regulation. To address this research gap, this study proposes an automatic knowledge graph construction approach based on deep neural network models. In total, 63 groups comprising 189 college students participated in this study and were assigned to three conditions, namely, the automatic activated and unactivated knowledge graph (AAUKG) condition, the automated activated knowledge graph (AAKG) condition, and the traditional online collaborative learning (TOCL) condition. The findings revealed that the AAUKG approach had more significant and positive impacts on collaborative knowledge building, group performance, social interaction, and socially shared regulation than the AAKG and TOCL approaches. This study provides substantial evidence of utilising both automatically activated and unactivated knowledge graphs to improve collaborative learning performance.