Semantic Preservation and Natural Language Data Augmentation via Variational Autoencoder

초록

Text augmentation, unlike image augmentation, is challenging because text modifications directly affect labels. Studies on text augmentation using generative and pretrained language models (PLMs) have been conducted; however, their application has limitations. This study proposes a PLM-based data augmentation technique using a variational autoencoder (VAE) structure. Latent variables were used to better understand the semantics, and the VAE was used to assign randomness. The PLM was placed in the encoder and decoder to improve the augmentation performance. We evaluated our proposed method on two benchmark datasets and demonstrated its augmentation effect.

키워드