Common Framework for Textual Anomaly Detection – CFTextAD

CFTextAD is a fundamental research project, contributing to axis E.2, "Artificial intelligence and data science". It proposes to build a common framework for textual anomaly detection, in order to understand how to best associate existing NLP representational models and existing anomaly detection methods. Hence, we propose to study, in particular, the transfer of knowledge from pre-trained models, and to characterize detected anomalies, by diversifying evaluation methods and focusing on interpretable methods. We intend to follow a line of research as general as possible, detaching ourselves from the classification setting underlying most of the current supervision and evaluation in anomaly detection, in order to obtain a tool allowing to improve the state-of-the-art for numerous natural language processing tasks.