Development of the Naïve Algorithm of Support System for the Hate Speech Detection in Social Media platforms

Abstract

While social media platforms provide a prominent space for users to participate in interpersonal conversations and express their viewpoints, the anonymity and façade afforded by these platforms may enable users to disseminate hate speech and objectionable information. Due to the extensive size of these platforms, there is a need to automatically detect and mark occurrences of hate speech. While there are several approaches available for detecting hate speech, the majority of these methods are intentionally designed to be non-interpretable or unexplainable. In this work, we want to overcome the problem of not being able to understand the results clearly. To do this, we suggest using advanced Large Language Models (LLMs) to extract certain characteristics from the input text. These characteristics, called rationales, will be used to train a basic hate speech classifier. This approach will ensure that the results are easily understandable and accurate. Our system successfully integrates the linguistic comprehension powers of LLMs with the discerning strength of cutting-edge hate speech classifiers to ensure that these classifiers are accurately interpretable. Our thorough assessment of many English language social media hate speech datasets reveals two key findings: (1) the effectiveness of the LLM-extracted rationales, and (2) the unexpected preservation of detector performance despite training for interpretability. Given the current rapid increase and exponential expansion of social media use, it is crucial to thoroughly examine social media information to identify any instances of hostile content. Researchers have been rigorously studying the differentiation between information that encourages hate and stuff that does not over the last decade. We propose a method to determine whether a speech fosters hatred or not by using both auditory and textual representations. The foundation of our technique relies on the Transformer architecture, which integrates audio and text sampling. Additionally, we have developed a unique layer dubbed "Attentive Fusion". Our research yielded superior outcomes compared to prior cutting-edge approaches, reaching an outstanding macro F1 score of 0.927 on the Test Set