Technology promises individualized teaching

Woolf addresses a growing need in education

Beverly WoolfEducation is no longer perceived as a “one size fits all” proposition.  Rather, the focus is on tailoring activities to meet the individual needs of an increasingly heterogeneous student population. This seems obvious, yet is nearly impossible to implement in a traditional classroom.

Even though teachers are responsible for more students in the classroom and online, a real possibility exists to produce a private computer tutor for each student, augmenting the teacher’s ability to respond to each individual.

Research Professor Beverly Park Woolf is building Web-based intelligent tutors that understand a student’s learning needs, optimize teaching materials and use effective tutoring strategies. “These tutors, when used in traditional classrooms with caring teachers, provide the advantages of individualized instruction at an affordable cost,” says Woolf.

Machine tutors operate like a trusted mentor, speeding through topics that the student grasps easily, concentrating on topics that cause trouble, and never losing patience. Some research challenges include:

Behavioral Studies and Student Modeling

Intelligent tutors are predicated on the existence of models of knowledge that represent the key ideas to be learned, common misconceptions, and how a student’s knowledge changes over time. Modeling an individual’s knowledge is challenging since students’ knowledge is often confounded or missing,.

Dr. Woolf has built many intelligent tutors in collaboration with colleagues in psychology, chemistry, engineering, biology, ecology, geology, education and medicine.

Photo: Geometry Tutor
Figure 1: The geometry tutor presents animation hints to mentor students through hundreds of geometry problems

Some of the most promising tutors were built with Research Scientist Ivon Arroyo and include tutors for elementary and high school mathematics (see Figure 1). The effects on individual learning differences (e.g., mathematical ability) and group characteristics (e.g., gender) were documented by deploying two tutors — AnimalWatch (arithmetic) and Wayang Outpost (geometry) — among nearly a thousand elementary and high school students. Large-scale experiments determined the practical significance of each tutor and showed that they have a measurable impact: students who used these tutors for two or three hours improved their results on Massachusetts-required standardized exams by 10-12%. 

These multimedia tutors contain nearly 1000 problems and are supplemented with data about cognitive features of each student, including variables for individual differences. For example, machine learning was used to modify tutor behavior according to each student’s Piagetian developmental stage, spatial ability, or math-fact-retrieval skills.

Evaluation results indicate that students with low cognitive skills learn best with concrete representations and manipulatives, and those with higher cognitive skills learn best with abstract or symbolic representations.  The effects of gender characteristics in learning have also been measured. Female students spend about 25% more time on hints than male students, perceive a tutor more positively than male students, and are more willing to use the tutor again. Boys with low cognitive development perform worse when they receive abstract or symbolic help while boys with advanced cognitive skills seem to learn better with abstract help than with hints.

In the private context of the tutor environment, students with weak skills benefit the most, seem comfortable requesting hints, make use of help and instruction and demonstrate improved performance. This is the reverse of the usual findings in the classroom, where higher achieving students are most likely to request help.

Learning to Teach

Machine learning (ML) techniques are used to model each student’s skills and to optimize the selection of problems and hints. Hierarchical, probabilistic models enhance the tutor’s ability to assess a student’s skills. The tutor reasons about latent variables, such as prior knowledge and the level of a student’s engagement in the tutoring process. Bayesian and data mining techniques help identify a student’s skills and predict student reactions to a variety of teaching styles (e.g., present a hint or an example) and to understand how each student learns.

Bayesian nets are used to reason about a student’s affective state (motivation, engagement, interest and learning) and to discover links between observable behavior (time spent on hints, number of hints selected) and hidden variables (attitudes and goals). Correlation between observable student activity and survey responses are converted into a network that tests the predictions on the data log of new students.

“Using ML techniques, we can predict a student’s level of engagement with 80-90% accuracy, and how a student will perform on each problem with 70% accuracy,” says Woolf.

Web-enabled intelligent instruction

Woolf is developing several tutors to personalize Web instruction based on presumed student knowledge, cognitive skills, and learning needs. One intelligent tutor can learn in just a few Web pages how to classify a student’s learning needs and which teaching approach to use, on a per-student basis. This tutor adapted each slide of a Web-enabled lecture course based on its prediction about which features a student would most like to see (e.g., definition, explanation, example). The tutor examined each topic at several difficulty levels and used a Naïve Bayes Classifier algorithm to determine whether or not the topic should be taught and to manage the layout of each page.

Another system generated an ontology of Web teaching materials to connect students with appropriate materials.

Beyond traditional classrooms

Students are often passive in classrooms; they are not regularly involved in thinking, active learning, problem solving or argumentation. In the traditional classroom, teachers ask 95% of the questions, mostly requiring short answers. Traditional classroom methods — lectures, books, multiple-choice exams — lead to passive students and are successful only with the top 25% of students. Liberal use of interactive graphics (3D modeling and interactive character animation) and sound within intelligent tutors help teachers connect with all students.

Woolf says “We do not intend for this technology to be used to imitate conventional classroom approaches; rather we focus on challenging traditional teaching and supporting new teaching methods.” Intelligent tutors play an essential role in moving education towards more student-centered methods, e.g., team collaboration, case-based inquiry and apprenticeship; techniques that are nearly impossible to implement without technology.

Photo: Inquiry Tutor
Figure 2: The Inquiry Tutor invites students to solve a medical diagnosis case (top) or to predict the location of the next earthquake (bottom).

The Inquiry Tutor immerses a student in cases that require critical thinking. For example, the student is invited to diagnose a medical disease or predict when the next earthquake will occur (Figure 2). A patient’s complaints form an initial set of data for the medical case and the student begins the process by ‘interviewing’ and ‘examining’ the patient. Students generate hypotheses which they then defend or refute with data accessed on the Web. Each investigation (e.g., questions, hypotheses, data collection and inferences) is tracked and students must articulate how evidence and theories are related.

“As cognitive science and psychology continue to broaden our understanding of how people learn, a real possibility exists to produce a teacher for every student,” says Woolf. Thus, content, teaching, assessment, student-teacher relationships and even the concept of educational institutions may all need to be rethought.

Woolf has an undergraduate degree in Physics from Smith College, received her Ph.D. in Computer Science from UMass Amherst and a second doctorate in Education. She joined the UMass Amherst Department of Computer Science in 1985 and is now Research Professor and Director of the Center for Knowledge Communication (ccbit.cs.umass.edu/ckc). Woolf is a Fellow of the American Association of Artificial Intelligence.