Hate Speech Scientific Project

Goal:

• To be able to recognize parts of text that contains hate or vulgarisms.

Possible applications:

• Management of discussion forums / detection of spam or abuse.
• "Postprocessing" for biased generative language models - preventing to generate inapropriate responses.

Plan:

• Perform a review of the state-of-the-art
• Pick established (english) corpora
• Formalize the problem - classification of sentiment, recognition of topic, keyword selection,
• Propose a preliminary system, repeat existing approach.
• Create small evaluation set in Slovak
• Try multilingual/crosslingual approach. Possibility of machine translation.
• Annotate a bigger Slovak Corpus
• Recognize and publish scientific contribution

Future Tasks:

• Evaluate existing multilingual model. E.G. https://huggingface.co/Andrazp/multilingual-hate-speech-robacofi
• Translate existing English dataset into Slovak. Use OPUS English Slovak Marian NMT model. Train Slovak munolingual model.
• Train or finetune or prompt a large langauge model.

In progress tasks:

• Annotate a Twitter Dataset. Possible guidelines are: https://developers.perspectiveapi.com/s/about-the-api-training-data?language=en_US
• Annotate a Facebook Dataset. Use some other guidelines. e.g. sentence-level annotation, for context sensitive hate.
• Prepare existing Slovak Twitter dataaset, train evaluate a model.

Finished tasks:

• Perform preliminary experiments with HS detection (Bulburu)
• Prepare an anotation infrastructure for Facebook data annotation (Ferko)
• Gather Facebook data and prepare for annotation. (Ferko)

People:

• Ján Staš
• Daniel Hládek
• Zuzana Sokolová
• Vladimír Ferko
• Tetiana Mohorian
• Patrik Pokrivčák

Former participants:

• Sevval Bulburu
• Manohar Gowdru Shridharu

Links:

• https://europeanonlinehatelab.com/
• https://hatespeechdata.com/
• https://oznacuj-dezinfo.kinit.sk/