Generative adversarial networks (GANs) are now one of the key techniques for detecting anomalies in images, yielding remarkable results. Applying similar methods to discrete structures, such as text sequences, is still largely an unknown. In this work, we introduce a new GAN-based text anomaly detection method, called ARAE-AnoGAN, that trains an adversarially regularized autoencoder (ARAE) to reconstruct normal sentences and detects anomalies via a combined anomaly score based on the building blocks of ARAE. Finally, we present experimental results demonstrating the effectiveness of ARAE-AnoGAN and other deep learning methods in text anomaly detection.
Tec Yan Yap, "Text Anomaly Detection with ARAE-AnoGAN" (2020). Honors Projects. 22.