Collaborative problem solving in computer-supported environments is of critical importance to the modern workforce. Coworkers or collaborators must be able to co-create and navigate a shared problem space using discourse and non-verbal cues. Analyzing this discourse can give insights into how consensus is reached and can estimate the depth of their understanding of the problem. This study uses Coh-Metrix, a natural language processing tool that measures cohesion, to analyze participant discourse from a recent multi-modal learning analytics study where novice programmers collaborated to use a block-based programming language to instruct a robot on how to solve a series of mazes. We significantly correlated thirty-five Coh-Metrix indices from the transcripts of dyads' discourse with collaboration, learning gains, and multimodal sensor values. We then fit a variety of machine learning classifiers to predict collaboration using the indices generated by Coh-Metrix as features. This study paves the way for real-time detection of (un)productive interactions from multimodal data and could lead to real-time interventions to support collaborative learning. [For the full proceedings, see ED599096.]