Understanding AdversarialBiasMitigator in more detail

[alicesaunders]

I was wondering how the adversarial bias mitigator worked in more detail. For example, in the config file there is the section of 'bias direction' (copied below). How does this tie in with the adversary and predictor models? Is this the gender information that the adversary is then trying to recover? Is there other information that is fed to the adversary elsewhere (i.e. from the training data) that is done within the training loop for the predictor/adversary?

I am attempting to mitigate gender bias from spanBERT for the task of coreference resolution. Is it sufficient to keep the bias direction below as it is? Or is it also necessary to add additional information about the protected variable elsewhere (either in the training loops or elsewhere in the config files)?

Thanks!

"bias_direction": {
"type": "two_means",
"seed_word_pairs_file": "https://raw.githubusercontent.com/tolga-b/debiaswe/4c3fa843ffff45115c43fe112d4283c91d225c09/data/definitional_pairs.json",
"tokenizer": {
"type": "pretrained_transformer",
"model_name": transformer_model,
"max_length": 512
}