University of Pittsburgh

Unsupervised Question Difficulty Modeling based on IRT, AFM and KT Models

ISP Graduate Student
Date: 
Friday, April 5, 2013 - 1:00pm

PLEASE NOTE:  THIS TALK WILL BE IN THE MARTIN COLLOQUIUM ROOM - 4TH FLOOR SENNOTT SQUARE

Question difficulty is useful in educational systems for suggesting students the suitable questions to solve, or suggesting teachers how to design suitable question sets.  Previous researches mainly  treat difficulty as an objective measurement and pay little attention to adapt to different students. Here, we define question difficulty as a subjective measurement considering a question's conceptual complexity (reflected somehow by current student's cognitive information) as well as the structural complexity (reflected somehow by group of students' cognitive information and the question's static information). Since there is no manually labeled difficulty in our dataset, two frameworks have been put forward: (1) select variables reflecting difficulty and then examine the relationship between these variables and our predicted difficulty (2) define different automatic labeling methods and examine the relationship between these labeled difficulty and our predicted values. For building the prediction model, we designed different ways to make use of current popular student modeling models (IRT, AFM, and KT), focusing on how to incorporate time factor, how to keep a balance between  individual differences and group of individuals' similarity.  We evaluate/compare them based on the above two frameworks.  Experimental results showed the soundness of our modeling method but also revealed some problems on which we need to further investigate.

Copyright 2009 | Web site by UMC Web Team