Text data augmentation using generative adversarial networks – a systematic review

Kanishka Silva, Kanishka Silva, Can, Burcu, Sarwar, Raheem ORCID: https://orcid.org/0000-0002-0640-807X, Blain, Frederic and Mitkov, Ruslan (2023) Text data augmentation using generative adversarial networks – a systematic review. Journal of Computational and Applied Linguistics, 1. pp. 6-38. ISSN 2815-4967

Preview

Published Version Available under License In Copyright. Download (266kB) | Preview

Abstract

Insufficient data is one of the main drawbacks in natural language processing tasks, and the most prevalent solution is to collect a decent amount of data that will be enough for the optimisation of the model. However, recent research directions are strategically moving towards increasing training examples due to the nature of the data-hungry neural models. Data augmentation is an emerging area that aims to ensure the diversity of data without attempting to collect new data exclusively to boost a model’s performance. Limitations in data augmentation, especially for textual data, are mainly due to the nature of language data, which is precisely discrete. Generative Adversarial Networks (GANs) were initially introduced for computer vision applications, aiming to generate highly realistic images by learning the image representations. Recent research has focused on using GANs for text generation and augmentation. This systematic review aims to present the theoretical background of GANs and their use for text augmentation alongside a systematic review of recent textual data augmentation applications such as sentiment analysis, low resource language generation, hate speech detection and fraud review analysis. Further, a notion of challenges in current research and future directions of GAN-based text augmentation are discussed in this paper to pave the way for researchers especially working on low-text resources.