Enhancing online safety: leveraging large language models for community moderation in Singlish dialect

Abstract

Online forums and comment sections have become ubiquitous features on social media platforms, providing users with the freedom to share their thoughts and opinions openly. This freedom, however, comes with a risk as users may abuse it by posting toxic or harmful comments. The anonymity and lack of immediate consequences on these platforms often embolden individuals to engage in disrespectful or offensive behavior, contributing to the proliferation of toxic content. In recent years, there has been a surge in interest surrounding transformers, largely driven by the success of models like ChatGPT. These transformer-based models have demonstrated remarkable capabilities in natural language processing tasks, including text generation and comprehension. As a result, researchers and practitioners have increasingly turned to transformers to address various challenges in the field of language processing, including content moderation. This project aims to leverage transformer technology to tackle the issue of toxic content in online forums, with a particular focus on Singlish forums. Singlish refers to a variety of English spoken in Singapore, characterized by its unique vocabulary, grammar, and intonation. The project seeks to fine-tune a pre-trained BERT model specializing in toxic comments detection, developed by Unitary, for content moderation of Singlish forums using Parameter Efficient Fine-Tuning (PEFT) methods.