Repository files navigation

Online grooming approaches, where a sexualpredator approaches minors online with thegoal of sexual abuse, are a big problem intoday’s world of social media. In this work,we present two approaches to detect sexualpredators in chats. We utilize the currentlyavailable datasets for Sexual Predator Detection (SPD) and analyze their strengths andweaknesses critically. Using dictionary-basedand transformer-based approaches, we analyzethe writing styles of predators in comparisonto non-predators in order to shed light totheir differences. Finally, we present ourtwo approaches, one of which improves thecurrent state-of-the-art score by 7.7%. Bothapproaches are based on BERT models usingadditional features of the chats as inputs.

Paper.pdf contains the documentation of our approach.Poster.pdf is a poster versionsrc/ is all the code. data/ is missing because the private dataset we used required access permission.

This project was developed in the context of the course "Computational Semantics for NLP" at ETH Zurich.My team Saahiti Prayaga, Philippe Schläpfer and I attempted to use NLP to tackle a pressing societal problem.