Includes updates and/or revisions.
Still, no one expects the technology to ever replace the real thing.
Then she responds encouragingly, “Let’s read again what the problem is asking.”
This scene would not be out of place in most educational environments were it not for one important detail: Jane is a character on a computer screen.
Developed by teams of researchers at the University of Massachusetts Amherst and Arizona State University in Tempe, the computer tutor in this scenario is part of a growing number of research projects around the country looking to build a social and emotional support system into intelligent-tutoring systems. Powered by artificial-intelligence technology, intelligent tutors have been around for decades, and some are beginning to make their way into regular classroom use. , one of the oldest and best known of such programs, for instance, now operates in 2,600 schools around the country.
But programming the systems to detect and respond to students’ emotions in the same way that human tutors do is a new wrinkle for the field and one that developers hope will enhance the technology’s educational potential.
More Resources
“Human tutors are believed to be the highest form of instruction,” said James C. Lester, a computer scientist who is exploring such ideas at North Carolina State University in Raleigh. “Why? The hypothesis is that there is very strong affective feedback going on. Five years ago, this was something the field would’ve thought was far out, but now it’s getting a lot more resources thrown at it.”
Major funders of research into imbuing digital tutors with the ability to detect and respond to students’ emotions include the National Science Foundation and the U.S. Department of Education’s Institute of Education Sciences.
The jury is still out on whether students learn more with emotion-enhanced computer tutors than they do with the other kinds of computer tutors, but some early results from small experiments around the country show promise.
In developed by the University of Massachusetts and Arizona State, which has been tested so far in a high school and in two schools of education, students’ passing rates on state tests were 10 percent higher after a week of lessons with the tutor than they were for peers who spent the same amount of time learning geometry in a regular classroom, according to Beverly Park Woolf, a research professor in computer science at the University of Massachusetts Amherst and a leading researcher on the project.
Her collaborators include Ivon M. Arroyo, also a research scientist at the University of Massachusetts, and Winslow Burleson, an assistant professor in the arts, media, and engineering program at Arizona State.
Their project is built around an intelligent-tutoring system known as Wayang Outpost, which uses real-world problem-solving tasks to teach geometry. The system picks up on students’ emotional states through hundreds of sensors embedded in the computer, students’ chairs, and other aspects of the students’ learning environment.
Sensors worn like a bracelet around the wrist detect changes in students’ pulse and in moisture levels on the surface of the skin. Sensors embedded in the chair cushions identify nine different postures a learner might take. Leaning forward, for example, suggests engagement, while a student leaning to the side might be bored or frustrated. On the computer mouse, pressure-sensitive sensors signal whether a student is squeezing harder in possible frustration.
The researchers also collect data on students’ emotions through a video camera embedded on the computer. It trains its focus on the student’s eyebrows, mouth, and nose, discerning whether the learner is smiling, frowning, or yawning.
In combination, the researchers find, the data collected through the sensors enables the system to correctly identify variation in students’ emotions more than half the time.
The program responds to students’ emotions in the form of “pedagogical agents”—which are animated characters, such as Jane—who mirror those emotions and offer an appropriate response. Building on the work of Carol S. Dweck, the Stanford University psychologist who argues that children work harder when they believe that intelligence is malleable rather than a fixed, inborn trait, the tutors also give feedback that emphasizes the value of effort.
“Hey, congratulations!,” Jane or “Jake” or “Isabel” might say, “Your effort paid off. You got it right!”
What the researchers have found in some of the studies so far is that the emotion-sensitive tutors, besides boosting students’ average achievement, seem to lead to improvements in the way that students think about math and their own math abilities. They are more likely after their tutoring sessions to agree, for instance, with statements saying “mathematics is an important topic” and to believe they are good at it. 69ý also seem to get bored with the tutor less quickly than they do with the unenhanced system.
Confidence Builder
Not all students liked the tutor, though. 69ý can switch off the animated tutors and boys, impatient to move along, often did. But girls, who started out in the studies expressing lower levels of confidence in their abilities than the boys did, also ended up feeling more confident than the boys when they used the tutor. The system also seems to be particularly effective with students who are classified to be in need of special education, Ms. Woolf said.
Ms. Woolf and her research partners also have found that boys and girls respond differently to pedagogical agents of a different gender. Researchers are now exploring what happens when the computer characters reflect the cultural orientations of their own racial and ethnic group.
“American Indians, for example, have a culture that talks about collaboration and not sticking out, which is opposite from the way most classrooms work, so the character can talk to American Indian students more about collaboration in the classroom,” said Ms. Woolf.
“In 20 years,” she said, “I think we will have personalized tutoring in the classroom.”
But other researchers, such as Arthur C. Graesser and Sidney D’Mello of the University of Memphis, have not found that students respond much differently to characters of different gender or racial or ethnic groups.
“We think it’s more what is said at the right time than what the agent looks like,” said Mr. Graesser, a psychology professor, “but that may not be true if students have to stay with a certain agent for weeks or months.”
In his work, Mr. Graesser has also found that the emotionally enhanced systems work better for poorer learners than they do for high achievers—but only after about half an hour.
‘Shake Up’ Comments
“You don’t want a tutor to get emotional too quick,” he said. “69ý seem to need conventional action to build up some bonding.”
Mr. Graesser and Mr. D’Mello experimented with injecting a dose of personality into their tutors.
“We called it a ‘shake up’ tutor,” he said. “If a student is frustrated, the tutor says, ‘I can see you’re frustrated. I thought you could handle this but I guess I need to rethink that.’ ”
But he said that approach hasn’t yet “panned out” in experimentation.
“We thought it might be more engaging for a certain class of students—maybe college students—who are a little more sophisticated.”
While the systems can be cumbersome for students to use and expensive to develop at first, all the experts said, they are becoming less so all the time.
“The sensors were very cumbersome, but now they’re just kind of cumbersome,” said Mr. Lester of North Carolina State, where researchers are integrating emotional feedback into game-based learning environments.
Within five to 10 years, he and others predict, costs may be low enough to make widespread use practical. Still, no one expects the technology to ever replace the real thing.
“There’s no way that I envision a class of 30 children going through the entire day on little machines,” Ms. Woolf said. “This will augment teaching. But as these systems become trustworthy, I absolutely think there will be a place in the classroom for them.”
