In recent years, the rapid development of artificial intelligence has increased the power of personalized learning. This study aimed to provide personalized intervention for each group participating in computer-supported collaborative learning. The personalized intervention adopted a deep neural network model, Bidirectional Encoder Representations from Transformers (BERT), to automatically classify online discussion transcripts and provide classification results in real time. Personalized feedback and recommendations were automatically generated from the classification results. A quasi-experimental research design was adopted to examine the effects of the proposed personalized intervention approach on collaborative knowledge building, group performance, socially shared metacognitive regulation, and cognitive load. Sixty-six college students participated in this study and were randomly assigned to the experimental and control groups. For online collaborative learning, students in the experimental group adopted the personalized intervention approach, whereas those in the control group used the conventional approach. Both quantitative and qualitative research methods were adopted to analyze data. The results indicated significant differences in the level of collaborative knowledge building and group performance between the experimental and control groups. Furthermore, the experimental group demonstrated more socially shared metacognitive regulation than the control group. There was no significant difference in cognitive load between the experimental and control groups. The results obtained from interviews were consistent with the quantitative data. The main findings together with the implications for practitioners are discussed in depth.