We explored the possibility of predicting student emotions (boredom, flow/engagement, confusion, and frustration) by analyzing the text of student and tutor dialogues during interactions with an Intelligent Tutoring System (ITS) with conversational dialogues. After completing a learning session with the tutor, student emotions were judged by the students themselves (self-judgments), untrained peers, and trained judges. Transcripts from the tutorial dialogues were analyzed with four methods that included 1) identifying direct expressions of affect, 2) aligning the semantic content of student responses to affective terms, 3) identifying psychological and linguistic terms that are predictive of affect, and 4) assessing cohesion relationships that might reveal student affect. Models constructed by regressing the proportional occurrence of each emotion on textual features derived from these methods yielded large effects (R2 = 38%) for the psychological, linguistic, and cohesion-based methods, but not the direct expression and semantic alignment methods. We discuss the theoretical, methodological, and applied implications of our findings toward text-based emotion detection during tutoring.