Arithmetic Problem Solver Based on BERT Model and Mathematical Cognitive Pattern

Abstract

Problem solving is a hot research area in educational technology. Arithmetic problem is the basis of many other problems thus it is important to develop algorithms to solve arithmetic problems. This paper proposes an arithmetic problem solver based on BERT model and mathematical cognitive pattern for solving the simple arithmetic word problems. Firstly, BERT model was used to obtain the semantic representation of the problem text. Then two different neural networks were adopted to identify the cognitive pattern and the variable contained in it successively. Finally, for each problem, all the variables are mapped to the corresponding cognitive pattern so as to instantiate an equation, and the program directly calculated the equation and got the answer. The experiment used 392 AddSub type problems from previous research and 474 Mix type problems collected in this paper. The result shows the solver achieved the 91.1% problems on Add-Sub type dataset and 75.5% problems on Mix-type dataset.