Data Augmentation with Large Language Models For Math Word Problem Solving

In this paper, we propose a method to enhance the problem-solving capabilities of models for Math Word Problems (MWP) using data augmentation with large language models. We aim to implement this method effectively; thus, three specific techniques are introduced: Data Augmentation Method Based on Probability Threshold, Iterative Correction-Based Problem Equation Generation, and Augmented Data Screening Based on Scoring Rules. These methods are applied at different stages before, during, and after the data augmentation process to achieve the best results. Overall, our approach works as follows: the model’s problem-solving ability is tested on the validation set by a certain step during the training, in order to collect both the problem types which the model struggles with and the instances which the model fails to solve. This data is then augmented using a large language model. After appropriate processing, this augmented data is integrated into the training set for subsequent training. Experimental results show that our method demonstrates superior performance on various datasets, with accuracy exceeding that of the baseline.