When Less Is More: Focused Pruning of Knowledge Bases to Improve Recognition of Student Conversation

Expert knowledge bases are effective tools for providing a domain model from which intelligent, individualized support can be offered. This is even true for noisy data such as that gathered from activities involving illdefined domains and collaboration. We attempt to automatically detect the subject of free-text collaborative input by matching students’ messages to an expert knowledge base. In particular, we describe experiments that analyze the effect of pruning a knowledge base to the nodes most relevant to current students’ tasks on the algorithm’s ability to identify the content of student chat. We discover a tradeoff. By constraining a knowledge base to its most relevant nodes, the algorithm detects student chat topics with more confidence, at the expense of overall accuracy. We suggest this trade-off be manipulated to best fit the intended use of the matching scheme in an intelligent tutor.