Fitting Software to Students

The use of ITS and other high-tech learning tools is increasing across the nation, but the effects are often below expectations. “Intelligent tutoring systems can provide effective instruction,” writes ISI researcher Carole Beal in a paper that will be presented July 20 at the AAAI 21st National Conference on Artificial Intelligence in Boston, “but learners do not always use such systems effectively.”

According to Beal, motivated students interested in course material take to ITS readily, but others will improvise ways to get through without putting in much effort: answering at random, or, quite commonly, abusing the program’s help feature by always asking for help as a way to get the answer without understanding the method.

Limiting access to the help function, for example, effectively defeats this last strategy — but doing so would hinder other students, for whom help is part of the learning experience.

To try to find out which students were most likely to game the system, Beal studied the behavior of a sample of 91 high school students working with a math ITS. Her method integrated three data sources: Students' reports on their own motivation; teachers' reports on the same students’ motivation; and, finally machine records of how the students in question used a web-based high school math tutoring system.

This last consisted of records of how students attacked math problems, and five different patterns emerged. Two of these were clearly unproductive. In one, students clearly selected answers at random, and kept doing so until they found the right answer by chance. In the other, they just started clicking on the help icon immediately after the problem was presented and kept clicking it repeatedly, to push through to the answer, and then repeating the process.

Matching up records with ITS behavior, some correlations were completely unsurprising. Students whose teachers identified them as motivated and who described themselves as motivated to do well in math showed little or no game- the-system behavior.

Other results were less obvious. “Proportionally speaking,” Beal reported, “students who described themselves as not good at math, not attracted to math, and not expecting to do well in math were most likely to use the ITS in a way that suggested a genuine effort to learn, by spending time reading the problem, and looking at the help features carefully and thoroughly”.

“The relatively high rate of learning-oriented ITS use by disengaged students suggests that technology-based instruction has potential to reach students who are not doing well with regular classroom instruction…. The opportunity to learn from software may offer an appealing alternative because the student can seek help in private.”

But between these poles, a large uncertain area remains. The largest single group of students was those with average motivation. About half of these followed learning strategies, the other half guessed. And the guessers were just as likely to be students whose teachers identified them as having higher math skills.

Within this group, however, one clue emerged. In the questionnaire used to elicit the self-descriptions, those who believed that mathematical skill was intrinsic, something students either had or didn't have, were more likely to guess. Those who thought math skill was something learnable were less likely to.

“This work is only a beginning,” says Beal. Her next step will be to use recently developed, sophisticated models of learning based on studies of expert human tutor, who (as Beal writes) accomplish their work “through a repertoire of feedback messages, sophisticated problem selection, and judicious offers of learner control when the learner appears to be flagging.”

By refining the ability to determine how a student is using the system — what their strategy is — Beal believes she and her team will be able to make ITSs more useful not just for the two categories of students using game-the-system strategies, but also for the other three, who seem to be trying to learn.

Beal's collaborators included graduate students Lei Qu and Hyokyeong Lee, both in the USC Viterbi School of Engineering computer science department; the work was funded by a grant from the NSF. Beal also holds an appointment as a research professor at USC's Daniel J. Epstein Department of Industrial & Systems Engineering.