This Machine Learning tool helps measure and visualize gender bias in word embeddings.
The tool is based on a pipeline that uses BERT embeddings as a starting point. The pipeline is composed by the following modules:
- Logistic Regression - l1 regularization
- PCA
- Support Vector Classifier
The tool is able provide a visualization instrument to analyze how much word is biased towards male or female gender, here is an example of the visualization:
Performance Metrics:
| Metric | Value |
|---|---|
| C (Logistic Regression) | 0.175 |
| C (SVM) | 0.375 |
| Accuracy | 0.7786 |
| Number of Selected Features (pre-PCA) | 38 |