Collaborative reflection (co-reflection) plays a vital role in collaborative knowledge construction and behavior shared regulation. Although the mixed effect of online co-reflection was reported in the literature, few studies have comprehensively examined both individual and group factors and their relationships that affect the co-reflection level. Therefore, this study explored the structural relationships between achievement goals (task-based, self-based, and other-based goals), online community identification, and co-reflection, which can consequently assist instructors in improving the related pedagogical strategies. To this end, 26813 posts on MOOC and college online learning platforms were gathered. Specifically, deep learning techniques were first used to train a classifier that classifies the large-scale co-reflection text automatically. The Bayesian method was then applied to disclose the structural relationships among achievement goals, community identification, and co-reflection. The results showed that the proposed classification algorithm achieved the best performance. Two best-fit models for characterizing the respective relationships between co-reflection and community identification as well as achievement goals were obtained using the Bayesian method. The results of the experiments on these two models demonstrated that both task-avoidance and other-avoidance goals were related directly to co-reflection, all task-approach, self-approach and other-approach goals were related indirectly to co-reflection, but self-avoidance goals had both a direct and an indirect relationship with co-reflection. The relationship between community identification and co-reflection was mediated by other-based goals. Some theoretical and practical implications were discussed for instructors and practitioners to build an online community.