Developing automatic hate speech detection methods based on text processing

Traditionally, sensory perception and hate speech detection have been viewed as distinct areas within the field of natural language processing (NLP). However, the profound impact of sensory experiences on cognitive and linguistic processes is undeniable. Previous NLP research has often overlooked the potential of integrating sensory knowledge with advanced language models for hate speech detection. Addressing this oversight, we introduce SensoryT5, a neuro-cognitive model that embeds sensory information into the T5 (Text-to-Text Transfer Transformer) framework, specifically tailored for enhanced hate speech detection. This innovative approach enriches the T5's attention mechanism with sensory cues, fostering a nuanced balance between contextual insights and sensory awareness. Through extensive testing across various datasets, SensoryT5 demonstrates superior performance in identifying hate speech, outperforming both the foundational T5 model and other pre-trained language models (PLMs)/large language models (LLMs) as well as traditional machine learning methods. This advancement not only boosts the model’s efficacy but also underscores the importance of neuro-cognitive data in augmenting the linguistic and emotional sensitivity of machine learning models. Our findings mark a significant evolution in NLP research, advocating for a more integrated approach that considers both cognitive science and computational techniques in tackling complex tasks such as hate speech detection.