Early Detection of Sexual Predators in Chats

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4985–4999 August 1–6, 2021. ©2021 Association for Computational Linguistics 4985 Early Detection of Sexual Predators in Chats Matthias Vogt Ulf Leser Alan Akbik Humboldt Universit¨at zu Berlin matthias.vogt@campus.tu-berlin.de leser@informatik.hu-berlin.de alan.akbik@hu-berlin.de Abstract An important risk that children face today is online grooming, where a so-called sexual predator establishes an emotional connection with a minor online with the objective of sex- ual abuse. Prior work has sought to automati- cally identify grooming chats, but only after an incidence has already happened in the context of legal prosecution. In this work, we instead investigate this problem from the point of view of prevention. We define and study the task of early sexual predator detection (eSPD) in chats, where the goal is to analyze a running chat from its beginning and predict grooming attempts as early and as accurately as possible. We survey existing datasets and their limita- tions regarding eSPD, and create a new dataset called PANC for more realistic evaluations. We present strong baselines built on BERT that also reach state-of-the-art results for conven- tional SPD. Finally, we consider coping with limited computational resources, as real-life applications require eSPD on mobile devices. 1 Introduction Online grooming denotes the process where a so- called sexual predator establishes an emotional con- nection with a minor online to systematically so- licit and exploit them for sexual purposes (Wachs et al., 2012). Online grooming is a major con- cern of public safety that, sadly, is rapidly growing. For instance, in England and Wales in the year to mid-2020, police recorded 5,083 offenses of Sex- ual Communication with a Child [1], an average of 14 offenses per day. In Germany, there were 2,632 recorded cases in 2020 where a child was sexually abused through internet communication technologies [2], an increase of 50 % to the previ- ous year. As such crimes often go unreported or undetected, police-recorded incidents certainly do ... ... Figure 1: Visualization of chat messages and early sexual predator detection (eSPD). On each new mes- sage, the analysis is updated to reflect the level of risk. Finally, an alert is triggered as the risk-threshold is passed. Our goal is to detect such risk as early as pos- sible. Note that real chats are much longer and can be non-contiguous conversations that span over weeks or months. Original source [3] not fully reflect the real scale of the issue (Bowles and Keller, 2019; McGuire and Dowling, 2013). The problem of detecting whether or not a child is being groomed by a predator is called sex- ual predator detection (SPD). Most previous ap- proaches to SPD have cast this as the problem of identifying predatory authors in a corpus of segments of chats (Villatoro-Tello et al., 2012; Cardei and Rebedea, 2017). Other approaches in- terpreted it as a binary classification problem over segments of a chat (Ebrahimi et al., 2016), or the entire …

Content truncated to 500 words