Projektbeschreibung
Filter für Hassreden auf ressourcenarme Regionen ausweiten
Im Zeitalter der sozialen Medien sind Hassreden zu einem alarmierenden globalen Problem geworden. Big-Tech-Unternehmen setzen maschinelles Lernen ein, um solche Inhalte zu filtern, doch für die meisten Sprachen mangelt es an Trainingsdaten. Finanziert über den Europäischen Forschungsrat werden im Projekt Respond2Hate mehrsprachige Repräsentationsmodelle entwickelt. So soll eine benutzerfreundliche Browsererweiterung geschaffen werden, mit der Personen hasserfüllte Inhalte unabhängig und lokal aus ihren sozialen Medien entfernen können. Dabei geht es konkret um flexible, adaptive Modelle, für die anfänglich nur wenige Trainingsdaten erforderlich sind. Daher werden modernste Verfahren zur Verarbeitung natürlicher Sprache und des Deep Learning eingesetzt. Indem Filter für Hassreden auf ressourcenarme Regionen ausgeweitet werden, die von Regierungen und NRO häufig wenig Beachtung erhalten, werden über dieses Projekt Nutzende befähigt, ihre Online-Erfahrung selbst zu gestalten.
Ziel
"Hate speech is a worldwide phenomenon that is increasingly pervading online spaces, creating an unsafe environment for users. While tech companies address this problem by server-side filtering using machine learning models trained on large datasets, these automatic methods cannot be applied to most languages due to lack of available training data.
Based on recent results of the PI's ERC project on multilingual representation models in low-resource settings, Respond2Hate aims at developing a pilot browser extension that allows users to locally remove hateful content from their social media feeds themselves, without having to rely on the support of tech companies.
Since hate speech is highly dependent on cultural context, responsive classifiers are needed that adapt to the individual environment. Commercial efforts focus on large-scale, general-purpose models which are often burdened with representation and bias problems, and therefore cope poorly with swiftly changing targets or information shift between regional contexts. In contrast, we seek to develop lightweight, adaptive models that require only a small dataset for initial fine-tuning by continuously enhancing model capabilities over time.
This is achieved by applying state-of-the-art Natural Language Processing (NLP) and deep learning techniques for pre-trained language models like low-resource transfer of hate speech representations from high-resource languages and few-shot learning based on limited user feedback. We have already successfully applied these methods in low-resource multilingual settings, and will now validate their use for hate speech filtering.
By making hate speech detection and reduction available in ""low-resource"" countries with little representation in current training datasets, which are currently not served well by governments, industry and NGOs, Respond2Hate will empower users to self-control their exposure to hate speech, fostering a healthier and safer online environment."
Programm/Programme
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Thema/Themen
Aufforderung zur Vorschlagseinreichung
(öffnet in neuem Fenster) ERC-2022-POC2
Andere Projekte für diesen Aufruf anzeigenFinanzierungsplan
HORIZON-ERC-POC -Gastgebende Einrichtung
80539 MUNCHEN
Deutschland