A Trustworthy Classifiers Chain for Cyberbullying Detection Using Bystander Roles

Cyberbullying has become a pressing societal challenge in the era of digital communication. Cyberbullying language is often part of a longer conversation that involves bystander roles. While incorporating additional contextual information—such as bystander replies—offers a more realistic and effective strategy for identifying instances of cyberbullying, most current studies are limited by their reliance on interpreting standalone Tweets. In this work, we explore fine-grained cyberbullying detection using bystander information through a chained model. However, sequential classification in chained models is prone to error propagation, where errors in earlier layers negatively affect the training and predictions in subsequent layers. Studies primarily focus on optimizing the chain classifier by developing strategies for selecting the optimal labels ordering. To improve the chained model with the pre-specified order as in our approach, it is crucial to enhance the reliability of the earlier layers to mitigate error propagation. This paper addresses this issue by incorporating a filter layer that determines the optimal threshold based on confidence and permissible error. This approach ensures that only predictions above the specified threshold are considered confident and carried over for use in the next layer of the chain. The results demonstrate that, using the Confident subset to train the upper classifier in the chain offers a distinct advantage over the standard chain model.