Mitigating Overfitting in Deep Learning Models for Machine Comprehension through Regularization and Data Augmentation

Abstract

Machine comprehension involves training models to understand and answer questions about given text passages, making it pivotal for applications ranging from automated customer service to expert systems. Despite considerable progress in deep learning, overfitting continues to pose significant challenges when models memorize training data rather than learning generalizable features. This issue becomes particularly evident in high-stakes settings, such as biomedical text interpretation or legal document analysis, where robust accuracy is essential. To address overfitting in deep learning models for machine comprehension, researchers have increasingly leveraged strategies such as regularization and data augmentation to promote better model generalization. Methods like dropout, weight decay, and batch normalization have contributed to reducing reliance on spurious correlations, while augmentation techniques further expand datasets to capture linguistic diversity and domain variability. Within the broader research community, there is a growing consensus that these two approaches—careful regularization and strategic augmentation—are among the most promising ways to mitigate overfitting. Still, an integrated understanding of how to optimally design, combine, and scale these practices remains limited. This paper investigates various regularization and data augmentation techniques, analyzes their effectiveness, and examines how they may be systematically integrated to enhance machine comprehension performance. In doing so, it seeks to provide a rigorous foundation for next-generation models capable of robust reasoning across diverse textual domains.