Improving Prediction Accuracy for Document Evaluation Problems Using Mixed-based Data Augmentation

Abstract

Data augmentation techniques are an essential part of computer vision and can provide significant accuracy gains at a small engineering cost. Inspired by Mixup, one of the data enhancement techniques for blending images, we applied sentence-by-sentence Mixup to text. We show that this improves the accuracy of the task of predicting English learners' writing scores compared to methods that do not use mixup.