Gender Biases Unexpectedly Fluctuate in the Pre-training Stage of Masked Language Models

Masked language models pick up gender biases during pre-training. Such biases are usually attributed to a certain model architecture and its pre-training corpora, with the implicit assumption that other variations in the pre-training process, such as the choices of the random seed or the stopping point, have no effect on the biases measured. However, we show that severe fluctuations exist at the fundamental level of individual templates, invalidating the assumption. Further against the intuition of how humans acquire biases, these fluctuations are not correlated with the certainty of the predicted pronouns or the profession frequencies in pre-training corpora. We release our code and data to benefit future research.