The School of Computer Science is pleased to present…
Osprey: A Reference Framework for Online Grooming Detection via Conversation Features and Backtranslation Augmentation
MSc Thesis Proposal by: Hamed Waezi
Date: Friday, 24 May 2024
Time: 10:00 am
Location: Essex Hall, Room 122
Abstract:
Online grooming is the process of an adult initiating a sexual relationship with a minor through online conversation platforms. To ameliorate the detection of such incidents, neural networks have been excessively employed, yet the models’ practical implications in real-world settings remain moot for their closed, shaky, irreproducible, and even poor evaluation methodologies under the sparse distribution of grooming conversations in the training datasets, like undermining recall over precision. Furthermore, proposed neural models overlook characteristic features of grooming in online conversations, including the number of participants (one-on-one), message exchange patterns (lack of turn-taking), and temporal signals, such as the elapsed times between messages. In this research, we work on the development of an open and standardized methodological framework aimed at augmenting reproducibility alongside an exploration of diverse feature-handling strategies evaluated against different metrics. Given the extreme class imbalance prevalent in our dataset, we further endeavor to explore the efficacy of roundtrip translation (backtranslation) and its impacts on the task of online grooming detection.
Keywords: Online Grooming Detection, Classification, Natural Language Processing, Translation
Thesis Committee:
Internal Reader: Dr. Robin Gras
External Reader: Dr. Tanja Collet
Advisor: Dr. Hossein Fani
