What is IBM’s Watson Education Initiative?
The way I like to look at our initiative is, we want to take advantage of all the interactivity that's exploding in computing and use them to structure learning experiences that are deeply immersive and deeply engaging. So the cognitive science part comes where we try to figure out what's actually happening in the brain as a function of all these emerging workloads and what does effective learning look like.
So the Signature initiative that we have is to use a computer to teach people using what we call an intelligent tutoring system, where one can access an intelligent tutor and works with the person in a very personalized fashion. Our global initiative essentially works on a whole series of such technologies. They are all deeply infused with AI techniques, like machine learning and natural language processing. There's a little bit of cognitive science and a flavor of cognitive science behind what we do.
How is IBM using Watson’s intelligent tutoring system?
So we are attempting to mimic the best practices of human tutoring. The gold standard will always remain one on one human to human tutoring. The whole idea here is an intelligent tutoring system as a computing system that works autonomously with learners, so there is no human intervention. It's basically pretending to be the teacher itself and it's working with the learner. What we're attempting to do is we're attempting to basically put conversational systems, systems that understand human conversation and dialogue, and we're trying to build a system that, in a very natural way, interacts with people through conversation. The system basically has the ability to ask questions, to answer questions, to know who you are and where you are in your learning journey, what you're struggling with, what you're strong on and it will personalize its pedagogy to you.
We have a fully conversation-enabled chatbot. A chatbot is kind of doing it a little bit less justice than I'd hoped but it's a fully conversation-enabled tutoring system that's working with students in college and in lifelong vocational courses. It's attempting to teach them their own topics that they're trying to master at their own pace and it's there as a non-judgmental 24/7 tutor that they can ask all kinds of questions to and it will also answer all the questions that they have.
We're doing this in a partnership with Pearson, which we formed late last year and we announced. We hope to cover a variety of disciplines that are taught in four year colleges, anywhere from psychology all the way to environmental sciences and physics and astronomy by the time we're done.
What makes Watson so effective at tutoring learners?
At the heart of our approach is an ability for computers to semantically understand language. It's a semantic understanding but it's not a human level understanding. There's a fairly deep understanding of what kinds of questions you're asking. For instance, you can phrase a question in any number of ways, the way people do, and the system still understands what the gist of the question is and it will retrieve the right answer for you. Similarly, it can ask you questions and you will provide your own responses in your own language and it will figure out if you've covered what it was looking for in an answer.
This whole exchange is also deeply personalized. It will figure out what the right question to service to you is based on what you've actually done so far in that particular course. So to address the former question, there's basically two major components behind it. There's a natural language understanding system and a machine learning system that's trying to figure out where you are in your learning journey and what the appropriate intervention is for you. The natural language system enables this interaction that's very rich and conversation-based, where you can basically have a human-like conversation with it and, to a large extent, it will try to understand and to retrieve the right things for you.
Again the most important thing is that we will set the expectations appropriately and we have appropriate exit criteria for when the system doesn't actually understand what you're trying to do.
What role does analytics play with Watson?
Learner models are basically a whole bunch of machine learning models where we're taking a lot of data. Depending on how instrumented the learning experience is, we can spend time looking at, for instance, how many times have we attempted a question, how many times have you tried to watch a video, how many times have you rewound the video, how long have you spent on a learning object. All of these are indicators to us about your state of confidence or mastery over a particular domain. Then, based on that, the system is turning around and recommending something else to do for you. The [inaudible 00:05:17] thing is basically analytics driven. It's all data driven. The evidence is data driven and it's based on what we think has worked in the past. At the same time, we're also marrying this with the rich ability to understand where the learner's at through their own language.