Comparative Analysis of Anomaly Detection Algorithms in Text Data

Proceedings of Recent Advances in Natural Language Processing, pages 1234–1245 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_131 1234 Comparative Analysis of Anomaly Detection Algorithms in Text Data Yizhou Xu1,2, Kata G´abor1, J´erˆome Milleret2, Fr´ed´erique Segond1,3 1ERTIM, INaLCO, 2 Rue de Lille, 75007 Paris, France 2ChapsVision, 4 rue du Port Aux Vins, 92150 Suresnes, France 3Inria, 860 Rue Saint Priest, 34095 Montpellier, France {yxu, jmilleret}@chapsvision.com kata.gabor@inalco.fr, frederique.segond@inria.fr Abstract Text anomaly detection (TAD) is a crucial task that aims to identify texts that deviate signif- icantly from the norm within a corpus. De- spite its importance in various domains, TAD remains relatively underexplored in natural lan- guage processing. This article presents a sys- tematic evaluation of 22 TAD algorithms on 17 corpora using multiple text representations, in- cluding monolingual and multilingual SBERT. The performance of the algorithms is compared based on three criteria: degree of supervision, theoretical basis, and architecture used. The results demonstrate that semi-supervised meth- ods utilizing weak labels outperform both un- supervised methods and semi-supervised meth- ods using only negative samples for training. Additionally, we explore the application of TAD techniques in hate speech detection. The results provide valuable insights for future TAD research and guide the selection of suitable al- gorithms for detecting text anomalies in differ- ent contexts. 1 Introduction Anomaly detection is a fundamental process in data analysis, aiming to identify inconsistent data points that deviate significantly from expected behaviors or established norms within a dataset. Such anoma- lies can emerge from various factors, including human errors, malicious behaviors, unusual events, or unexpected changes. Effective anomaly detec- tion can facilitate swift problem recognition, proac- tive measures for error correction, and future prob- lem prevention. It enhances data quality, aids in risk identification, and empowers decision-making across diverse domains, with its utility extending to various data types such as tabular data, graphs, time series, texts, images, and videos. In the context of text data, the anomalies refer to specific texts or textual fragments that deviate significantly from established norms, which can be determined based on the overall text or cor- pus, regular language usage, or common sense. These anomalies may manifest at various linguis- tic levels, such as orthographic (spelling), lexical (word usage), syntactic (sentence structure), se- mantic (meaning), and discourse (overall context) levels (Wang et al., 2014; Saranya et al., 2014; Wahl, 2021; Sufi and Alsulami, 2021). Detecting anomalies in text data holds vital importance in ap- plications like language development assessment, plagiarism detection, quality control in data pro- cessing, and identifying abnormal language usage in cybersecurity (Cichosz, 2020; Szopl´ak and An- drejkov´a, 2021). In this article, we concentrate on Text Anomaly Detection (TAD) at the semantic and discourse lev- els, where norms are established on a corpus scale. It is important to note that the definition of anoma- lies can be further refined and may slightly differ according to specific contexts. For instance, in the realm of competitive intelligence, anomalies often relate to abnormal themes or topics, while in the field of online reputation monitoring, they typically …

Content truncated to 500 words