On Measuring and Mitigating Biased Inferences of Word Embeddings
Abstract
Word embeddings carry stereotypical connotations from the text they are trained on, which can lead to invalid inferences in downstream models that rely on them. We use this observation to design a mechanism for measuring stereotypes using the task of natural language inference. We demonstrate a reduction in invalid inferences via bias mitigation strategies on static word embeddings (GloVe). Further, we show that for gender bias, these techniques extend to contextualized embeddings when applied selectively only to the static components of contextualized embeddings (ELMo, BERT).
Cite
Text
Dev et al. "On Measuring and Mitigating Biased Inferences of Word Embeddings." AAAI Conference on Artificial Intelligence, 2020. doi:10.1609/AAAI.V34I05.6267Markdown
[Dev et al. "On Measuring and Mitigating Biased Inferences of Word Embeddings." AAAI Conference on Artificial Intelligence, 2020.](https://mlanthology.org/aaai/2020/dev2020aaai-measuring/) doi:10.1609/AAAI.V34I05.6267BibTeX
@inproceedings{dev2020aaai-measuring,
title = {{On Measuring and Mitigating Biased Inferences of Word Embeddings}},
author = {Dev, Sunipa and Li, Tao and Phillips, Jeff M. and Srikumar, Vivek},
booktitle = {AAAI Conference on Artificial Intelligence},
year = {2020},
pages = {7659-7666},
doi = {10.1609/AAAI.V34I05.6267},
url = {https://mlanthology.org/aaai/2020/dev2020aaai-measuring/}
}