Keita Kurita

Measuring Bias in Contextualized Word Representations

2019

Contextual word embeddings such as BERT have achieved state of the art performance in numerous NLP tasks. Since they are optimized to capture the statistical properties of training data, they tend to pick up on …