Unleashing the Potential
by Alexander Gelfand
We’re going to beat Google at the game of ranking pages,” says Neil Heffernan, associate professor of computer science and co-director of the university’s new graduate program in learning sciences and technologies. That goal may sound ambitious, but Heffernan has reason to be optimistic.
He is the researcher behind ASSISTments, an intelligent, web-based tutoring system that can assess student performance, offer tailored help (the name is a combination of “assistance” and “assessments”), and track progress over time, capturing every detail of a student’s interaction with the software and generating detailed reports for teachers. Teachers can choose questions from existing content libraries or construct their own, and the system adapts to each student’s responses in real time, providing the specific kind of help — suggestions, hints, reminders — they most need.
Although it can handle just about any subject, ASSISTments has thus far been used primarily for math, and education officials at both the state and federal levels have shown interest. Over the past year, nearly 20,000 K–12 math students — most of them in Massachusetts and Maine — have used ASSISTments either in the classroom or to complete homework. Heffernan recently received a $500,000 grant from Next Generation Learning Challenges (funded by the Bill and Melinda Gates and the William and Flora Hewlett foundations) to double the number of middle school students who use the software, and he aims to increase that number to a million by 2015. The system supports several different constituents, including school administrators, who want better ways to conduct schoolwide tests; classroom teachers, who want to customize the instruction they offer; and students themselves, whose academic performance has improved as a result of using the system, according to a number of randomized controlled trials.
Heffernan and his graduate students are constantly seeking ways to improve and build upon the platform. For example, they recently applied for significant funding for a project that would allow schoolchildren in Arkansas to access ASSISTments using text messages, and they developed an ASSISTments iPhone app that is being used by students in Maine. Another new feature under development would recommend the most useful Web pages for students seeking to master one of the 134 specific math skills that the ASSISTments platform addresses (prompting the comparison to Google).
“We’re always experimenting,” he says. “We as a nation need proven educational interventions. If we do not know what works, then we have to experiment to find out what does.”
Visualizing Science Literacy
While the ASSISTments team continues to refine its platform, another WPI research team is using it as a base upon which to build a revolutionary approach to augmenting and assessing science education.
“I started with the ASSISTments system developed for math instruction,” says Janice Gobert, a cognitive scientist and associate professor of social science and policy studies, “and constructed on top of it a whole new infrastructure specifically designed for the quite different challenge of teaching and assessing scientific inquiry.”
Developing an intelligent system that can help teach middle school students to think like scientists is no trivial undertaking, she says. Put simply, doing science is not the same as doing math. Her group needed to develop a system that could guide students through the basic steps of scientific inquiry — formulating hypotheses, designing experiments, analyzing data, and communicating findings — and assess in real time just how well they’re doing at those inquiry skills and learning the concepts of middle school science.
Transformative science often comes from the intersection of two disciplines.— Janice Gobert
To do that, Gobert, who co-directs WPI’s Learning Sciences and Technologies Program with Heffernan, built upon her previous research on the role that visualization plays in science learning to develop a system based on interactive “microworlds”: rich visual simulations of natural phenomena culled from physics, life science, and earth science that obey the rules of the scientific phenomena they address. Using an on-screen widget, students can manipulate these microworlds to design and run their own experiments — exploring the relationship between weight and volume by switching between different kinds of liquids and containers, for example, or sussing out the effect of temperature and volume on phase changes by bringing different quantities of liquids to a boil.
Gobert’s system uses algorithms to analyze student inquiry strategies in real time and recommend alternatives. She is currently developing a virtual agent that combines principles derived from cognitive science with predictive data-mining techniques developed by her colleague Ryan Baker to pinpoint each student’s skills. (Because the science ASSISTments infrastructure is designed to record every interaction a student has with it, it is able to provide a wealth of data that researchers can sift for meaningful patterns.) The agent, a cartoon dinosaur named Rex, can ask struggling students questions and then provide targeted help — breaking complex tasks into simpler ones, recommending fresh strategies — to nudge them along.
Digging Deep for Answers
The computational power behind Rex is a cutting-edge example of educational data mining, which uses sophisticated algorithms to analyze the vast amounts of data collected by educational software. When combined with insights from cognitive science, it has proved to be an especially powerful tool. “Transformative science often comes from the intersection of two disciplines,” Gobert notes. Nowhere is this more evident than in the range of projects undertaken by Baker, assistant professor of psychology and the learning sciences, who brings expertise in computer science and educational psychology to his research.
For example, Baker, who was recently elected the inaugural president of the Educational Data Mining Society, has built automated detectors that can be embedded in educational software to determine when students are “gaming the system,” or trying to get the right answers without actually thinking through the material. He is working with Gobert to develop automated systems that use statistical and probabilistic techniques to not only assess whether a student has learned a scientific skill, such as how to design a controlled experiment, but to also predict whether that student will be able to use that skill in the future. And he and Gobert are working to study student engagement with science education, a subject that could have far-reaching implications for the nation’s schools and workforce.
I believe the combination of computer science and cognitive science will, over the next couple of decades, produce the same kind of revolution in education that they have already brought about in chemistry, physics, and medicine.— Ryan Baker
For example, they have studied the relationship between a student’s engagement with educational software and the quality of science learning he or she will do later on in the school year. He has partnered with Heffernan to investigate the relationship between student engagement with mathematics and the decision to pursue a career in math. And he is working with Albert Corbett, a scientist at Carnegie Mellon University who helped build the world’s most widely used intelligent tutoring system, to develop a detector that can predict whether a student will be prepared for future learning beyond the educational software.
Having software that can answer such fundamental questions as whether a student is engaged in learning and whether he is prepared to learn in the future will allow scientists to build intelligent tutoring systems that can intervene earlier and more effectively when students are struggling. Such systems might even prevent students from too hastily abandoning the pursuit of careers in math, science, engineering, and technology — a matter of crucial importance given the increasing need for citizens who are well-versed in these STEM disciplines and the disappointing number of graduates in those fields.
“Educational data mining allows us to figure out what’s working and what’s not working in a much more effective way than was ever possible before,” Baker says. “I believe the combination of computer science and cognitive science will, over the next couple of decades, produce the same kind of revolution in education that they have already brought about in chemistry, physics, and medicine.
“It’s great to be at WPI — a place that’s really at the vanguard of this emerging field.”