From Hate to Harmony: Leveraging Large Language Models for Safer Speech in Times of COVID-19 Crisis

Abstract

This study investigates the rampant spread of offensive and derogatory language during the COVID-19 pandemic and aims to mitigate it through machine learning. Employing advanced Large Language Models (LLMs), the research develops a sophisticated framework adept at detecting and transforming abusive and hateful speech. The project begins by meticulously compiling a dataset, focusing specifically on Chinese language abuse and hate speech. It incorporates an extensive list of 30 pandemic-related terms, significantly enriching the resources available for this type of research. A two-tier detection model is then introduced, achieving a remarkable accuracy of 94.42% in its first phase and an impressive 81.48% in the second. Furthermore, the study enhances paraphrasing efficiency by integrating generative AI techniques, primarily Large Language Models, with a Latent Dirichlet Allocation (LDA) topic model. This combination allows for a thorough analysis of language before and after modification. The results highlight the transformative power of these methods. They show that the rephrased statements not only reduce the initial hostility but also preserve the essential themes and meanings. This breakthrough offers users effective rephrasing suggestions to prevent the spread of hate speech, contributing to more positive and constructive public discourse.