Skip to content

Investigating Gender Bias in NER Models

Here you can find descriptions of the ner task in GenDa Lens.

Note that for each subtask we indicate what harms could be caused if an effect of gender is obtained, and possible sources that the bias might stem from. You can read more about these under User Guide/Defintions.

Idea Behind Framework

This framework quantifies bias as a systematic difference in error (error disparity). In the context of gender bias, with a binary understanding of gender, this means that if the error distribution for women is different than the error distribution for men the model is biased.

The framework relies on the concept of data augmentation in conjunction with DaNe, which is a Danish data set that can be used for finetuning NER models.

The DaNe data set is an extended version of the The Danish Universal Dependencies Treebank (UD-DDT). It is annottaed with named entitites for the folllowing four categories: persons (PER), organizations (ORG), locations (LOC) and miscellaneous (MISC) for each token.

Specifically, the idea behind the framework is to use data augmentation to replace all the PER entities in the DaNe test set with either male or female names and then evaluate how well the model performs NER on the test set.

Task

For the investigation of gender bias two augmentations of the DaNe test set are made:

  1. Replace PER entities with all Danish female names

  2. Replace PER entities with all Danish male names

NER performance is then evaluated with F1 scores.

Evaluation: Main Effect

Possible Harms: Possible Bias Sources:
Underrepresentation Selection Bias

The overall effect of gender on this task is based on F1 scores for the femala and male augmentation respectively.

Specifically it is calculated as:

Evaluation: Nuance

Possible Harms: Possible Bias Sources:
Underrepresentation Selection Bias

For the detailed bias evaluation it is assessed whether a possible gender bias in the model interacts with a bias towards minorities. This is what is referred to as intersectional bias in the literature (see e.g. Subramanian, S., Han, X., Baldwin, T., Cohn, T., & Frermann, L. (2021).). Specifically the following augmentations are made:

  1. Replace PER entities with all Danish female names

  2. Replace PER entities with all Danish male names

  3. Replace PER entities with all minority female names

  4. Replace PER entities with all minority male names

For this task performance is then calculated with F1 scores for each of these four augmented data sets.