FROM SENTIMENT TO SAFETY: CONDITIONAL HATE SPEECH DETECTION BASED ON EMOTION ANALYSIS IN SOCIAL ROBOTICS

This study explores the integration of emotion recognition and sentiment analysis with hate speech detection in the context of social robotics and natural language processing (NLP). The research is based on the assumption that affective cues-such as negative sentiment or specific emotional states like anger-can serve as effective triggers for selectively activating hate speech detection modules, enhancing both efficiency and contextual accuracy. The methodology involved three main stages: a review of existing hate speech datasets, the selection and evaluation of classification algorithms for sentiment and emotion analysis, and an experimental assessment of the relationship between affective features and offensive language. The study examined five datasets in Polish and English, emphasizing the cultural and contextual specificity of hate speech, which often includes political undertones and idiomatic expressions. Two sentiment analysis models and three emotion recognition models, including a proprietary solution, were compared. The experiments showed a strong alignment between negative sentiment and the presence of offensive content. Similarly, emotions such as anger and, in some cases, surprise, were frequently associated with hateful expressions, depending on the dataset and model used. These findings suggest that affective analysis can function as an effective preliminary filter for triggering hate speech detection in NLP systems.