Abstract

The effectiveness and predictive performance of text classification models depend significantly on the availability of large, diverse, and accurately labeled datasets, which enable machine learning algorithms to learn meaningful linguistic patterns and achieve robust generalization across different domains. However, acquiring high-quality labeled textual data is often expensive, time-consuming, and domain-dependent. Data augmentation has emerged as an effective solution to address data scarcity by generating additional training samples that improve model generalization. Among various augmentation techniques, autoencoders to learn latent semantic representations and generate meaningful synthetic text samples. This research presents a comprehensive study of data augmentation for text classification using autoencoders. The proposed framework employs an encoder-decoder architecture to generate semantically similar text representations that enrich training datasets and improve classification performance. The study discusses the theoretical foundations, implementation methodology, evaluation metrics, advantages, challenges, and future research opportunities associated with autoencoder-based text augmentation. Experimental observations indicate that incorporating augmented textual data enhances classification accuracy, reduces overfitting, and improves model robustness, particularly in low-resource and imbalanced dataset scenarios.