Using Automated Dialog Analysis to Assess Peer Tutoring and Trigger Effective Support

Intelligent tutors have the potential to be used in supporting learning from collaboration, but there are few results demonstrating their positive effects in this domain. One of the main challenges in automated support for collaboration is the machine classification of dialogue, giving the system an ability to know when and how to intervene. We have developed an automated detector of conceptual content that is used as a basis for providing adaptive prompts to peer tutors in high-school algebra. We conducted an after-school study with 61 participants where we compared this adaptive support to two nonadaptive support conditions, and found that adaptive prompts significantly increased conceptual help and peer tutor domain learning. The amount of conceptual help students gave, as determined by either human coding or machine classification, was predictive of learning. Thus, machine classification was effective both as a basis for feedback and predictor of success.