How can interactive systems model their users better by listening to them? Conversely, how can systems listen better by exploiting user models? We propose to investigate these questions in the context of a inter-disciplinary challenge problem: developing a computational student model of children's oral reading -- that is, a model of individual student knowledge, behavior, and learning that can guide a tutor. This work will integrate and extend methods from speech technology, cognitive psychology, user modeling, and intelligent tutors.
The proposed work will draw on, contribute to, and test and refine the scalability of an innovative, technology-enabled intervention and unique research platform developed with previous NSF support: an automated Reading Tutor that displays stories on a computer screen, uses a speech recognizer to listen to children read aloud, responds with spoken and graphical assistance, and helps them learn to read. Its ability to listen enables novel continuous assessments of students' reading progress. As of 2003, the Reading Tutor was used daily by hundreds of children on 216 computers in nine schools. The proposed work will use speech and other valuable data captured by the Reading Tutor.
Children who use the Reading Tutor have improved significantly more in reading comprehension and other skills than statistically matched controls. However, its current effectiveness is limited by inability to accurately hear and model the student.
The challenge is to analyze children's oral reading and estimate various component literacy skills at a sufficiently fine grain size to guide the decisions of an intelligent tutor, so as to adapt to students' individual or collective educational needs.
The challenge of developing a student model of oral reading will provide a problem-driven basis for guiding the work and quantifying its success. Incorporating successively better models in the Reading Tutor will both test their accuracy and put them to immediate use in improved literacy tutoring. Controlled studies will investigate how much successive versions of the Reading Tutor increase learning gains for different types of students. Automated experiments embedded in the Reading Tutor will help analyze which tutorial actions help in which cases.
The proposed research combines broad technical and direct societal impact. Expected technical contributions of this work include advances in speech recognition (especially for young and disfluent speakers), user modeling, intelligent tutoring, and automated assessment of comprehension and other literacy skills. Broader impacts include improved literacy for thousands of children who use the Reading Tutor during the study, and many more thereafter.
