hate-alert/Fear-speech-analysis

The project curates a dataset of fear speech from Indian WhatsApp political groups and builds classification models using doc2vec with SVM/logistic regression and fine-tuned BERT transformers to detect messages designed to incite fear rather than explicit toxicity. It includes network analysis notebooks examining user-user propagation patterns and textual analysis using topic modeling and emoji sentiment (Empath), revealing that fear speech spreads faster despite lower toxic signal than traditional hate speech classifiers detect. The repository integrates Hugging Face Transformers for NLP and uses privacy-preserving survey methodology to validate real-world impact on group members.

No commits in the last 6 months.