Using Artificial French Data to Understand the Emergence of Gender Bias in Transformer Language Models
This work addresses the problem of understanding gender bias emergence in AI models for researchers, though it is incremental as it builds on prior studies of linguistic property learning.
The study investigated how transformer language models learn gender information by training them on an artificial French corpus with controlled gender distributions, finding that models can capture gender correctly under certain conditions but exhibit bias in others.
Numerous studies have demonstrated the ability of neural language models to learn various linguistic properties without direct supervision. This work takes an initial step towards exploring the less researched topic of how neural models discover linguistic properties of words, such as gender, as well as the rules governing their usage. We propose to use an artificial corpus generated by a PCFG based on French to precisely control the gender distribution in the training data and determine under which conditions a model correctly captures gender information or, on the contrary, appears gender-biased.