Addressing religious hate online: from taxonomy creation to automated detection

Area of Science:

• Computational Linguistics
• Social Media Analysis
• Hate Speech Detection

Background:

• Online abusive language, particularly misogyny and racism, has been extensively studied.
• Religious hate speech remains underexplored due to scattered data and varying annotation schemes, hindering effective computational containment.
• Existing datasets lack interoperability and do not adequately address the country-specific nuances of religious hate speech.

Purpose of the Study:

• To propose a fine-grained, interoperable labeling scheme for religious hate speech detection.
• To introduce a novel, annotated Twitter dataset in English and Italian covering Judaism, Christianity, and Islam.
• To evaluate various classification algorithms and assess cross-lingual transferability for religious hate speech detection.

Main Methods:

• Development of a fine-grained taxonomy for abusive language, specifically tailored for religious hate speech.
• Creation and annotation of a Twitter dataset in English and Italian based on the proposed scheme.
• Experimentation with machine learning and transformer-based language models for classification tasks.
• Investigation of cross-lingual transferability using multilingual models.

Main Results:

• Demonstrated the effectiveness of the proposed labeling scheme and dataset for religious hate speech detection.
• Evaluated the performance of different classification algorithms, highlighting the challenges in detecting religious hate speech.
• Showcased the potential for cross-lingual transfer learning in low-resource scenarios.

Conclusions:

• The proposed fine-grained labeling scheme and dataset enhance the study of religious hate speech.
• The research provides valuable insights into computational approaches for detecting religious hate speech across languages.
• The findings support the development of more robust and widely applicable tools for online safety.