In the context of the final project for Le Wagon Data Science & AI course, we have built an NLP model able to detect sexism in text. The initial classification will be a binary 'sexist' or 'not sexist'.
As opposed to previous research and models, we used a dataset composed of 6 text corpi, not limited to social media.
- “Call me sexist, but...”, 2021
- Explainable Detection of Online Sexism 2023
- sEXism Identification in Social neTworks 2021
- Expert Annotated Dataset for the Detection of Online Misogyny 2021
- Sexist Stereotype Classification 2020
- Detection of Sexist Statements ... at the Workplace 2020
As the datasets use different annotation rules, we hoped to provide a richness and nuance to the model, at the risk of being 'overly accusatory'.
Our final model, which can be tested here, is a fine-tuned BERT model, trained for Precision.
The options for further exploration include, but are not limited to:
- Developing a french model
- Developing a multi-lingual model
- Augmenting our dataset(s)
- Translation
- Scraping reddit/instagram/youtube/twitch
- Generative AI
- Text templates (eg, "I hate women, they're all (bitch/slut/whore)s")
- Annotating our own data set based on gender theory and language theory
- Multi-class classification