Experiments in building Intelligent Tutoring Systems
I have been interested in how learning and thinking occur for almost as long as I have been a student. I find that when I diagnose my own performances (particularly failures) in assessments (quizzes/ back of the chapter problems), it comes down to the following:
When I was thinking of the answer to a problem:
1. I did not have the relevant facts in my long-term memory: lack of learning
2. I had the facts in my long-term memory, but my brain didn’t cue them as relevant to the problem, or the connections formed were not strong enough: lack of targeted and adequate practice
3. I had confidence in incorrect facts and used them to solve the problem: (misconceptions/assumptions)
4. Somewhere my reasoning was wrong.
These mistakes may occur in isolation, or together in series when it comes to higher level problems I faced at work such as designing a reactor, formulating a test plan while troubleshooting a complex customer problem, or brainstorming ideas for improvements to a product.
This is where the Socratic method of teaching becomes powerful. Forcing active participation (as opposed to passive listening in the case of classroom lectures) when information is disseminated is one way to ensure that the above problems are reduced. Although typically used with open ended questions, it is useful in examining the facts and rules we have accepted as true. With the right kind of questions from the teacher, it facilitates critical and creative thinking that is crucial to learning.
In the Socratic method, the teacher asks thought provoking questions, and the student answers them. The teacher then challenges the student with follow up questions, so the student self-assesses the accuracy of their assumptions and reasoning. This then forms the basis for the next natural question that needs to be answered. This is similar to the background intellectual process when a student is reading a text. The student processes each new fact in the text to see if it fits with his current understanding, and if it does, he stores it in his long-term memory, and if it does not fit, he grapples with the schema he has in his head until the contradictions are resolved.
Could a computer possibly play the role of Socrates for my personal learning goals? What about an individualized Socrates for every student in class? There would be many benefits of such an application. Below are some of the ideas
- Every student could learn at their own pace.
- The background knowledge of every student is different, so questions could be ordered/designed in a manner optimized to each student’s personal needs.
- Reviewing and practice is targeted and adequate. This will avoid working on several questions that deal with a small subset of the important concepts or being presented with a problem for which one is not quite ready yet.
- The computer would throw up contradictions between facts inferred from previously accepted assumptions and new facts the student comes across. The student could then resolve them right away instead of building further understanding on shaky foundations.
But Socrates is an expert. Does the computer need to be an expert to play Socrates? Could we create a poor man’s expert by extracting a concept map from text? Concept maps are graphs/diagrams that consist of labelled connections (relationships) between different concepts. If we could extract key terms from text, and their relationships with one another, we would be able to then construct a concept map to represent an ‘expert level understanding’.
Extracting concept maps from text is not a new idea. The field of computational linguistics and natural language processing has made significant leaps in the last few decades and there are several publications that showcase the ability to extract information from text. In this series, I describe my experiments and progress with building an Intelligent Tutoring System from STEM (science, technology, engineering and mathematics)
