In a previous study, we found that real-time mutual gaze perception (i.e., being able to see the gaze of your partner in real time on a computer screen while solving a learning task) had a positive effect on student collaboration and learning (Schneider & Pea, 2013). The goals of this paper are (1) to explore a variety of computational techniques for analyzing the transcripts of student discussions; (2) to examine whether any of those measures sheds new light on our previous results; and (3) to test whether those metrics have any predictive power regarding learning outcomes. Using various natural language processing algorithms, we found that linguistic coordination (i.e., the extent to which students mimic each other in terms of their grammatical structure) did not predict the quality of student collaboration or learning gains. However, we found that a simple computational measure of student verbal coherence (i.e., the extent to which students build on each other's ideas) was positively correlated with their learning gains. Additionally, this measure was significantly different across our experimental conditions: students who could see the gaze of their partner in real time were more likely to develop a coherent discussion. Finally, using various language metrics, we were able to roughly predict (i.e., using a median-split) learning gains with a 94.4% accuracy using Support Vector Machine. The accuracy dropped to 75% when we used our model on a validation set. We conclude by discussing the benefits of using computational techniques on educational datasets.