An Implicit Numbers and Pre-trained model based Math Word Problem Solver

• 1、School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：Math word problem solving is an interesting task that needs to bridge the natural language descriptions and formal mathematical equations. Previous research has attempted to introduce end-to-end neural network models that generate expression templates and then fill in the numerical values from the problem text into the expression. However, these models are powerless when it comes to numerical that exist in the mathematical expression but are not explicitly mentioned in the question text. To address this issue, this paper proposes a very effective method called the Implicit Common-Sense Number Set (ICNS) scheme, which can dynamically include implicit common-sense numbers into the context representation of mathematical expressions. The ICNS includes numbers that are not explicitly mentioned in the problem text but can be inferred from it and exist in the mathematical expression. Experimental results on the public datasetMath23K demonstrate that Our model achieves accuracies of 77.6%.

Keywords： Math Word Problem Pre-trained model Implicit Common-Sense Number Set