Obfuscation-aware Cyberbullying Detection using BERTweet with Conditional Leet-speak Normalization

Abstract

Cyberbullying through digital platforms is a major concern on social media. Pretrained transformer encoders have performed well on detection tasks involving clean social-media text; however, little research has examined their robustness when characters are substituted with visually similar digits or symbols. This study proposes an obfuscation-aware framework consisting of a transformer encoder pretrained on Twitter data and a leet-speak normalization pipeline. The rule-based detector checks whether each input contains leet-speak; if leet-speak is detected, a character-level normalizer processes the input before encoding; otherwise, the input is passed to the encoder unchanged. The framework was evaluated on two publicly available English cyberbullying datasets, each containing 15,592 instances. Six encoders from three pretraining domains were evaluated on clean and synthetically obfuscated data using three random seeds. The results show that all six encoders degraded under the evaluated synthetic leet-speak transformations, with drops of up to 48% points relative to their clean-text baselines. The proposed pipeline improves leet_test macro-F1 by 18.13% points on the first dataset and by 13.92% points on the second, without reducing performance on clean text. A heuristic lower-bound analysis was used to assess whether the gains remained above the baseline across the tested random seeds.