An end-to-end grading neural network for middle-school math problems
Meng Lin · Illinois Digital Environment for Access to Learning and Scholarship (University of Illinois at Urbana-Champaign) · 2019
Mathematics homework grading is a common real-world task where a human grader checks a student's solution to a math problem against the answer key and gives a score. This thesis proposes a deep-learning-powered grader that takes the place of the human grader. The task is formulated as a classification problem. Given an answer key and a student's solution, the classifier needs to predict two metrics: (1) a four-class classification result that measures the completeness of the student's detailed steps and (2) a binary classification result that identifies whether the conclusion of the student's solution is accurate. A new model, Step Comparison Transformer (SCT), is introduced, and its performance is validated on a set of grading data provided by a commercial provider of artificial intelligence products for education.