Collaborative game-based learning enables students to engage in rich problem-solving activities where they can work together to learn science concepts while developing critical collaboration skills. However, facilitation is often required to support effective collaborative learning processes, which requires teachers to provide personalized support to each group of students. Although a teacher can attend to issues that may arise during collaborative learning, such as out-of-domain chat messages, socioemotional tensions between group members, or challenges with completing the problem-solving activity, it is impractical for teachers to continuously monitor multiple groups at the same time to know when to intervene and provide guidance or assistance. In this paper, we introduce predictive models for facilitation intervention that analyze student collaborative chat dialogue as students engage in problem-solving in a collaborative game-based learning environment. We first collected a dataset of expert facilitators intervening in students’ collaborative problem-solving as groups of students engaged in collaborative learning in middle school science education focused on ecology. We then trained our predictive models on this collected facilitated student interaction data to predict whether the facilitator would intervene. The best-performing predictive model outperforms competitive baselines for predicting when a facilitator will provide assistance to students, paving the way toward robust facilitation support for collaborative game-based learning.