Comparing Zero-Shot Text Classification and Rule-Based Matching in Identifying Cyberbullying Behaviors on Social Media
Wei Jiek Chong, Hui Na Chua, May Fen Gan · 2022 IEEE International Conference on Artificial Intelligence in Engineering and Technology (IICAIET) · 2022
The incidences of cyberbullying have skyrocketed due to the continuous expansion of social media users. Therefore, proactive efforts are necessary to address cyberbullying, including countermeasures for handling the various cyberbullying behaviors. Nevertheless, due to the large volume of social media texts being generated persistently, it is challenging to identify cyberbullying behaviors in a social text and not scalable in using the manual approach of human annotation. Most previous studies adopt the human annotation approach to determine whether a text is a cyberbully or non-cyberbully. Therefore, this paper aims to experiment with approaches that can improve the efficiency of recognizing the different cyberbullying behaviors through textual data using zero-shot classification and rule-based matching, and compare how they perform in classifying cyberbullying behaviors. This study uses techniques such as topic modelling, zero-shot text classification, and information extraction with rule-based matching to identify and classify the cyberbully behaviors underlying a cyberbully comment. The human annotation approach serves as the benchmark to compare the performance of both models in identifying cyberbullying behaviors. Our results show that zero-shot classification performed better accuracy in categorizing cyberbullying behaviors. Among the behaviors, the zero-shot model we generated presents a better accuracy rate in recognizing the flaming behavior, but it achieves a lower accuracy rate in identifying the other cyberbully behaviors.