Reducing Bias in Cyberbullying Detection with Advanced LLMs and Transformer Models
Click the Poster to View Full Screen, Right click to save image
Grant: L.E.A.P Scholarship
Dahana Moz Ruiz
CoPIs:
Annaliese Watson, Anjana Manikandan, Zachary Gordon
College:
The Dorothy and George Hennings College of Science, Mathematics, and Technology
Major:
Computer Science
Faculty Research Advisor(s):
Yulia Kumar, J.Jenny Li, Patricia Morreale
Abstract:
This paper delves into a comprehensive exploration of the inherent biases present in Large Language Models (LLMs) and various Transformer models. These models, which are optimized for the purpose of identifying and dissecting instances of cyberbullying, form the focal point of our investigation. The aim is to refine and enhance the accuracy and fairness of these models by mitigating the deeply rooted biases that permeate their structures. This is essential to target because language models can inadvertently perpetuate and amplify existing biases present in the data they are trained on.
The foundation of this study is robust, relying on empirical data meticulously gathered from 'X' (former Twitter), where cyberbullying is systematically classified into several categories including Age, Ethnicity, Gender, Religion, Other and includes instances where no cyberbullying is observed. This allows for a comprehensive and well-rounded dataset for analysis. A sophisticated cyberbullying detection application was engineered, utilizing the advanced capabilities of the OpenAI API as its backbone. This application serves as a filtering mechanism, designed to issue alerts for content that is identified as inappropriate, thereby fostering a safer and more respectful online environment.
In this study, an overview of the latest innovations and developments in cyberbullying detection, emphasizing the pivotal roles of LLMs and Transformer models was presented. Our main research questions under study are: How does the implementation of bias mitigation strategies impact the accuracy and fairness of models in detecting cyberbullying across diverse demographic groups? Can LLMs like ChatGPT improve the cyberbullying detection compared to previously well-known Transformer neural networks? Can LLMs like ChatGPT ease and simplify the cyberbullying detection compared to previously well-known Transformer neural networks?
The aim is to uncover a deeper understanding of the multifaceted landscape of cyberbullying detection, thereby contributing to the ongoing efforts to create a safer and more inclusive digital space.