Artificial intelligence can make MOOCs smarter
Massive open online courses (MOOCs) have democratized higher learning by delivering free classes from stellar universities online to anyone, any place.
But its flaw is right there in the name – massive.
With no limits on class size, one course can have thousands of students, making individualized attention to students an overwhelming prospect to educators. Might this lack of attention deter students from finishing the course?
As of 2017, 78 million students worldwide have signed up for at least one course, according to data collected by Class Central. Since 2011, when Stanford University launched its first MOOC that kickstarted the MOOC movement, 9,400 courses have been offered from more than 800 universities.
However, the median rate of enrolled MOOC students who complete a course is just 12.6 percent, according to a 2015 report published by the International Review of Research in Open & Distributed Learning, leaving lots of room for improving retention.
Artificial intelligence (AI) presents opportunities to solve these issues. Those include customizing learning paths tailored to an individual students’ capabilities and creating opportunities for more teaching assistance – from humans and AI. According to a December 2017 study published in Nature, incorporating AI-based strategies would enhance MOOC students’ learning experiences.
To tailor a course, data mined from students of previous offerings of the MOOC can determine what profile students fit. “Then, the system knows how to classify them,” said Han Yu, assistant professor of computer science and engineering, Nanyang Technological University, Singapore, and co-author of the Nature article. “The previous class data serves as the baseline, but then their personal information customizes the course to the individual.”
“Differentiated learning programs will be used in all learning aspects in the future – whether in the classroom or online.”
Software with goal-oriented modeling, such as Goal Net, could adjust a student’s learning sequence by accounting for the skill set they already bring to the course. The pedagogy would not change but the delivery of content and the engagement with students would, Yu said.
“Customization is doable and possible,” said Greg Kessler, associate professor of Instructional Technology at Ohio University, Athens, Ohio. “We should anticipate that these differentiated learning programs will be used in all learning aspects in the future – whether in the classroom or online.”
Researchers at University of Melbourne developed a model to predict who is likely to drop out of a MOOC, based on information from two offerings of a Coursera course.
“We used some simple measurements, such as how many lectures they watched, how many documents they downloaded, and attempts made on assignments. It turns out you can make very accurate prediction of who will drop out fairly early on,” said James Bailey, professor of Computing and Information Systems, co-author of a 2015 study.
The effort, using transfer-learning algorithms, updates the students’ information each week, calibrating their chances of success over time. “In principle, you could alert students early on who have about a 50% chance of passing or failing and encourage them,” Bailey added.
Human and AI teaching assistants
Research shows that less than 10% of MOOC students use online forums, Nanyang Technological’s Yu said. Part of the problem is that instructors don’t have enough time to answer each individual. “Students then leave the forum and sometimes the class,” he added.
To offer students more individual attention, MOOCs need to recruit and retain more teaching assistants who previously took the course and thus can respond to the current students’ questions and online comments. Crowdsourcing techniques could help find community teaching assistants, Yu said.
Meanwhile, Georgia Institute of Technology launched its “Jill Watson,” an AI teaching assistant that responds to student introductions and questions, for a MOOC in January 2018.
“Our vision is to build a Jill Watson for every class, at every school, at every level.”
Originally, the AI teaching assistant (TA) was built to help an online AI class. Since 2016, Jill has been greeting students and answers their questions in discussion forums within three to 20 minutes for Georgia Tech’s online course Knowledge-Based Artificial Intelligence, a core course in its Master of Science in Computing Science program. Since Jill joined the team, student online engagement has increased from an average of 32 comments to an average of 38, according to the university.
Now, the virtual TA is being used for the first time in Introduction to Computing, a MOOC with 60,000 registered students around the world, said Ashok Goel, Georgia Tech professor of computer science and creator of Jill Watson. The MOOC course has no scheduled human interaction.
Jill’s use in a MOOC is a step toward making virtual learning companions universally available. “Our vision is to build a Jill Watson for every class, at every school, at every level,” Goel said. “It would be available to every student, teacher and parent.”
Jill’s sophistication has evolved over time. Georgia Tech has customized Jill’s responses to adjust to the student’s emotional state based on the wording of their posts. “We can tune the content and delivery to different categories,” he added.
MOOCs provide millions of pieces of data on each student’s interactions that could create richer learning experiences for students in all types of classrooms and improve learning across the board.
But, getting the humans in charge of the MOOCs to share that information with AI researchers remains a challenge,
“MOOCs provide millions of pieces of data on each student’s interactions that could create richer learning experiences.”
“MOOCs are very closed,” Yu said. “Data is not shared across the board, and everything depends on individual researchers working in cooperation with the MOOCs. It takes a lot of negotiation and that remains an issue.”
In China, companies like Alibaba and Didi regularly open up various types of large-scale de-identified datasets of consumer behaviors to discover good algorithms that can help their businesses, he said.
Of course, privacy remains a chief roadblock. But, according to Yu, data safeguards like those used in the private sector can be utilized to ensure that data being shared do not disclose personal information.