TAD-Bench: A Comprehensive Benchmark for Embedding-Based Text Anomaly Detection

Existing studies often lack Anomaly detection is a critical task in machine systematic evaluations of how different embeddings learning, with applications ranging from fraud detection perform across diverse anomaly types, raising and content moderation to user behavior questions about their generalization capabilities analysis (Pang et al., 2021). Within natural language in complex, real-world scenarios such as multilingual processing (NLP), anomaly detection has become settings or domain-specific anomalies. Recent increasingly relevant for identifying outliers efforts, such as AD-NLP (Bejan et al., 2023) such as harmful content, phishing attempts, and and NLP-ADBench (Li et al., 2024), have significantly spam reviews. However, while AD tasks in structured advanced anomaly detection in NLP. ADdata (e.g., tabular, time series, graphs) (Steinbuss NLP provides valuable insights into different types and Böhm, 2021; Blázquez-García et al., 2021; of anomalies, while NLP-ADBench expands evaluations Qiao et al., 2024) have achieved significant maturity, to a wide range of algorithms and datasets.