Unsupervised Anomaly Detection in Multi-Topic Short-Text Corpora

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1392–1403 May 2-6, 2023 ©2023 Association for Computational Linguistics Unsupervised Anomaly Detection in Multi-Topic Short-Text Corpora Mira Ait-Saada1,2 and Mohamed Nadif 1 1Centre Borelli, Université Paris Cité, 75006, Paris 2Caisse des Dépôts et Consignations, 75013, Paris {mira.ait-saada,mohamed.nadif}@u-paris.fr Abstract Unsupervised anomaly detection seeks to iden- tify deviant data samples in a dataset without using labels and constitutes a challenging task, particularly when the majority class is hetero- geneous. This paper addresses this topic for textual data and aims to determine whether a text sample is an outlier within a potentially multi-topic corpus. To this end, it is crucial to grasp the semantic aspects of words, particu- larly when dealing with short texts, since it is difficult to syntactically discriminate data sam- ples based only on a few words. Thereby we make use of word embeddings to represent each sample by a dense vector, efficiently capturing the underlying semantics. Then, we rely on the Mixture Model approach to detect which sam- ples deviate the most from the underlying distri- butions of the corpus. Experiments carried out on real datasets show the effectiveness of the proposed approach in comparison to state-of- the-art techniques both in terms of performance and time efficiency, especially when more than one topic is present in the corpus. 1 Introduction Anomaly Detection (AD) is a task that can address various objectives such as mining frauds (Deng and Mei, 2009), diseases (Han et al., 2021) and intru- sions (Pu et al., 2021). AD takes several forms: supervised, unsupervised or semi-supervised. Un- supervised AD implies that no prior information about the dataset is provided. In this case, the so- lution usually consists of identifying samples that deviate in a certain way from the others among the same dataset; anomalies being, by definition, rare phenomena. Particularly, anomalies in a tex- tual dataset can be defined as samples having an atypical vocabulary (lexical anomaly) or a devi- ating global meaning (semantic anomaly). Iden- tifying abnormalities in textual data can be very useful in many industrial use-cases. A good exam- ple is the detection of non-eligible and/or fraudu- Les lignes de commande Linux pour débutants inlier Formation: Introduction au Shell Bash inlier Administration système Unix pour les nuls inlier Apprendre à utiliser le terminal Ubuntu/Debian inlier Formation en Espagnol pour débutants outlier Table 1: Example illustrating the importance of seman- tic representations in a small corpus of short texts. lent course contents in the public French plateform MonCompteFormation1 where millions of course sessions are available with no possibility of control- ling training organizations in a supervised fashion (using labeled data). Hence, to assess the effec- tiveness of our approach, we rely on an external labeled dataset that closely relates to course con- tents and that is dedicated to course certifications. The dataset is described in Appendix A. In addition to the difficulty of mining anomalies in short-text corpora of varying sizes, we also have an important computational cost …

Content truncated to 500 words