Audio: preventhate.org, 16 October 2025

(Gemini Audio)

• According to scientific reports and Twitter data from major global debates, findings are that online discourse is primarily structured by shared ideological alignment rather than by prominent actors. Users form polarized communities across topics, with selective exposure and shared narratives reinforcing ideological divides.
• The review by MDPI synthesizes findings from 79 political science and international relations studies on the causes and consequences of online hate speech. It highlights key drivers such as platform policies and far-right rhetoric, and outlines counterstrategies including allyship, digital literacy, and deterrence mechanisms.
• Multi3Hate introduces a culturally annotated dataset of 300 memes across five languages via arXiv to assess how vision-language models handle hate speech. The study reveals significant cultural bias in model performance and annotation agreement, underscoring the need for culturally adaptive moderation tools.
• The dataset by Data in Brief offers 26,985 high-confidence Malay-English social media texts for binary hate speech classification, addressing the underrepresentation of Southeast Asian languages in NLP. It supports multilingual machine learning, cross-lingual benchmarking, and educational applications for English and Malay-speaking communities.
• The study published by arXiv proposes a Retrieval-Augmented Generation framework for counter-speech, leveraging a curated knowledge base of 32,792 authoritative texts. Evaluated on the MultiTarget-CONAN dataset, the framework outperforms existing baselines and enables scalable, reliable counter-speech generation for eight targeted groups.
• This arXiv-issued work compiles and standardizes hate speech datasets for European Spanish, Portuguese, and Galician variants to support multilingual detection. It establishes new benchmarks using large language models and emphasizes the importance of variety-aware approaches in low-resource language contexts.