This work analyse the gender bias of a Portuguese pre-trained Language Model and performs an investigation of different methods for modifying an embedding to remove gender stereotypes. In our experiemnts, we've found few examples of gender bias and stereotypes and demonstrated that is possible to reduce gender bias in Portuguese embeddings while preserving its properties such as analogy tasks.
Results and more information available on our preprint PDF
@article{santana2018there,
title={Is there Gender bias and stereotype in Portuguese Word Embeddings?},
author={Santana, Brenda Salenave and Woloszyn, Vinicius and Wives, Leandro Krug},
journal={arXiv preprint arXiv:1810.04528},
year={2018}
}