Deep learning techniques for hate speech detection

• Main Author: Sam, Jared Mun Kit
• Other Authors: Luu Anh Tuan
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Computer science and engineering::Computing methodologies::Document and text processing
• Online Access: https://hdl.handle.net/10356/172646
• Institution: Nanyang Technological University

Considering the prevalence of hate speech in social media platforms, automatic hate speech detection is a crucial tool in the fight against hate speech proliferation. Several techniques, such as the recent surge in deep learning-based methods, have been developed for the task. Different datasets that represent different facets of the hate speech detection issue have also been created. Using three prominent public datasets, a comprehensive empirical analysis of hate speech detection techniques is presented in this study. The implementation and comparison of current models offered pivotal insights into machine learning models’ efficacy, word representation models, and their performance variance across different datasets. Convolutional Neural Networks (CNN) emerged as a consistent performer, especially when coupled with Bidirectional Encoder Representations from Transformers (BERT) embeddings. The performance of Multi-Layer Perceptron (MLP) was notably affected by the chosen word representation method, with the BERT combination being superior. Word representation evaluation underscored BERT’s superior capability, attributable to its pre-training on extensive corpora and its provision of contextual word representations, outclassing fixed embeddings like Global Vectors for Word Representation (Glove) and Term-Frequency-Inverse Document Frequency (TF-IDF). Despite BERT’s strengths, its low macro average scores highlight the challenges in accurately identifying minority hateful tweets amidst vast tweet volumes.