Implications of Using Internet Sting Corpora to Approximate Underage Victims

Abstract

Law enforcement officers (LEOs) and the justice system employ NLP models for classifying and triaging child exploitation cases due to the textual communications between predators and victims. The usefulness of these systems depend on the quality of data that can be used for training. Data in the domain are scarce, sensitive, and emotionally taxing for annotators. NLP researchers approximate victimization conversations using transcripts from internet stings performed by either vigilantes or LEOs, with an implicit assumption that vigilante or LEO conversations represent the victimization process. Psychology research, however, states that underage victim chats differ from internet stings in goal and modus operandi. We present a methodology and observations from annotating a corpus of victim, vigilante, and LEO conversations with convicted predators with the goal of comparing these chats. The corpus is annotated for stages and tactics of the victimization process described within psychology research. As predicted by psychological research, we found significant differences in the three classes of chats that are usually not taken into account in chat classification.