ADVANCED LARGE LANGUAGE MODEL APPROACHES AND NATURAL LANGUAGE PROCESSING TECHNIQUES TO IMPROVE HATE SPEECH DETECTION USING AI

Abstract

The proliferation of hate speech on social networks can create a significant negative social effect, making the development of AI-based classifiers that can identify and characterize different types of hateful speech in messages highly important for stakeholders. While this is a highly challenging problem, recent advances in language models promise to advance the state of the art such that even subtle and indirect forms of hate speech can be detected.

In this dissertation we present a series of contributions that improve different aspects of hate speech classification. We developed THOS, a hate speech dataset consisting of 8.3k tweets. Compared to previous datasets, THOS contains fine-grained labels that identify not only whether a tweet is offensive or hateful, but also the target of the hate. Using this dataset, we studied the degree to which finer grained classification of tweets is feasible. In the follow-up work, we focus on the difficult problem of implicit hate speech, where hate is conveyed through subtle verbal constructs and allusions, without the use of explicitly offensive terms. We evaluate the efficacy of lexicon-based methods, transfer learning, and advanced LLMs such as GPT-4 on this problem. We found that the proposed techniques can boost the detection performance of implicit hate, although even advanced models often struggle with certain interpretations. In our third contribution, we introduce the Closest Positive Cluster (CPC) auxiliary loss, which improves the generalizability of classifiers across a wide range of datasets, resulting in enhanced performance for both explicit and implicit hate speech scenarios. Finally, given the scarcity of implicit hate speech datasets and the abundance of explicit hate datasets, we proposed an approach to generalize explicit hate datasets for the classification of implicit hate speech. Additionally, the proposed approach addresses noisy label correction issues commonly found in crowd-sourced datasets. Our method comprises three key components: influential sample identification, reannotation, and augmentation. We show that the approach improves generalization across datasets and enhances implicit hate classification.