Modularized Learning for Hate Speech Detection in Korean: Integrating Emotions and Multi-Faceted Attributes

Detecting hate speech is critical both online and offline, particularly on social media, in personal reviews, and during interpersonal conversations. In this study, we present a modularized learning algorithm for hate speech detection in Korean that integrates diverse features, including emotional indicators and attributes capturing expressions of hatred. To investigate the influence of complex human emotions on hate speech, we leverage multiple datasets for both emotion classification and hate speech detection. Our approach, which integrates these varied datasets, enhances performance and examines the impact of multi-label annotation on detection accuracy. Built on pre-trained Korean language models, our framework extracts features related to emotion, multiple attributes, and similarity, and then employs a classifier for final prediction. Experimental results demonstrate that our proposed method outperforms existing algorithms by achieving a higher f1-score in hate speech recognition.