Segmentation and Graph Matching for Online Analysis of Student Arithmetic Operations

This paper is based on a research project aiming at improving learning long arithmetic operations in primary school using pen-based tablets. The goal is to automatically analyze a student's handwritten answer by comparing it to an expected answer and to provide immediate feedback. This comes down to find any mistake made such as a calculus mistake, missing carry over or symbol misalignment. We use the correspondence obtained by the Graph Edit Distance (GED) computed between both the student and expected answers. In order to reduce graph sizes to overcome the computational complexity of the GED on large graphs, we present a new semantic graph of line segmentation. We propose a backtracking process to correct potential early mis-recognition mistakes for non-corresponding vertices. We evaluate the improvement on the analysis performances for an increasing number of backtracks on an in-house dataset composed of 400 handwritten operations.