Effective and efficient user and content modeling for intelligent tutoring systems

Suleyman Cetintas, Purdue University

Abstract

Most effective teaching methods actively engage students in the process of learning. Active engagement requires individual attention of teachers for each student group or student, which is highly time and resource intensive and almost impossible to implement in most schools in most disciplines. In the last four decades, many intelligent tutoring systems (ITS) have been developed in several domains to provide individualized guidance without a need for one teacher for each student, and have been shown to produce the same improvements as do the effective teaching methods. In order to be effective, intelligent tutoring systems require i) sufficient and high quality educational content, ii) detailed, domain specific content models to assess the difficulty levels of educational materials, iii) fine-grained, domain specific student models and additional equipment such as microphones, cameras, and sensors, and iv) an intelligent recommendation module that automatically provides students with interesting content of appropriate difficulty. However, it is very time-consuming and costly to prepare sufficient and high quality educational content as well as to build domain specific student and content models. They require intensive labor from the domain experts; therefore, have long been recognized as major bottlenecks for the development of ITS. Furthermore, additional equipment such as microphones and cameras that are used for modeling students behaviors outside the tutoring system are not available in most public schools. This dissertation studies novel probabilistic approaches for effective and efficient user and content modeling for intelligent tutoring systems. In particular, we propose methods that i) analyze and model (i.e., detect the types of as well as identify the relevant and irrelevant information in) educational content without requiring domain experts' help, ii) model students' off-task behaviors with the equipment available in most schools, iii) model students' performance without a need for domain expert knowledge, and iv) jointly model students and educational content for more effective student and content modeling. A fully-functioning prototype system has been developed and evaluated in local schools. Empirical studies conducted on real-world datasets from the prototype system as well as on external large-scale datasets demonstrate the effectiveness of the proposed models.

Degree

Ph.D.

Advisors

Si, Purdue University.

Subject Area

Artificial intelligence|Computer science

COinS