Textual anomaly detection: A systematic evaluation of representations and algorithms

This study evaluates scalable, unsupervised methods for detecting malicious online content by benchmarking transformer-based architectures across diverse linguistic contexts. We demonstrate that the semantic resolution of the embedding layer is the critical determinant of performance. Our results identify that large language models coupled with manifold learning achieve superior anomaly separation in knowledge-intensive domains, significantly outperforming traditional BERT-based pipelines. Additionally, we reveal a topological dichotomy in detection strategies: Contrastive autoencoders offer robust stability in structured environments, whereas few-shot deviation learning (FATE) is essential for high-entropy, dynamic contexts such as politics and sports. These findings propose a shift toward context-aware architectures capable of adapting to the complex semantic landscape of modern web content.