Collaborative activities, like peer tutoring, can be beneficial for student learning, but only when students are supported in interacting effectively. Constructing intelligent tutors for collaborating students may be an improvement over fixed forms of support that do not adapt to student behaviors. We have developed an intelligent tutor to improve the help that peer tutors give to peer tutees by encouraging them to explain tutee errors and to provide more conceptual help. The intelligent tutor must be able to classify the type of peer tutor utterance (is it next step help, error feedback, both, or neither?) and the quality (does it contain conceptual content?). We use two techniques to improve automated classification of student utterances: incorporating domain context, and incorporating students' self-classifications of their chat actions. The domain context and self-classifications together significantly improve classification of student dialogue over a baseline classifier for help type. Using domain features alone significantly improves classification over baseline for conceptual content.