Sexism detection with deep learning and LLMs
Bi-LSTM and Transformer classifiers versus prompted Mistral models on the EXIST dataset. The Transformer reached a test F1 of 0.83.
- Trained Bi-LSTM and Transformer classifiers on the EXIST dataset: test F1 of 0.71 and 0.83
- Evaluated Mistral V2 and V3 on the same tweets with zero-shot and few-shot prompting
- Mistral V2 reached 0.74 F1 zero-shot (with 30 of 300 prompt-following failures); Mistral V3 reached 0.72 three-shot with no failures
- Date
- Jan 2025
- Domain
- Deep learning & NLP
- Stack
- Bi-LSTM, Transformer, GloVe, Mistral