Gender Bias in Automated Email Responses

Generating Bias-Free Smart Email Responses for Microsoft

Industry Challenge

This project aimed at tackling gender bias in Microsoft’s Smart Reply feature. Microsoft’s current approach consisted of removing gendered responses and replacing them with gender-neutral variants semi-automatically. However, this often led to unnatural sounding responses in gendered languages such as French, Spanish, etc.

The researchers at ADAPT needed to come up with an approach that wasn’t too dependent on linguistic tools and could easily be applied to different gender-inflected languages.

The ADAPT Solution

This project aimed to develop a cost-effective, automated solution which could be scalable to a larger data set for Machine Learning model training. 

To do this, ADAPT researchers adopted an interdisciplinary study combining linguistic and technological knowledge to first identify sentence segments that should have the opposite gender variant. A gender variance was then created through a large gender-parallel Spanish corpus for those words using a rule-based approach. 

Finally, Python language was used to train a neural rewriter to automatically generate gender variance on unseen data sets without the need to pre-process, making the solution extremely scalable for gender-inflected languages.

Benefits/Impact

  1. Using technology to be a force for inclusion and non-discrimination. 
  2. Removal of data bias to tackle gender discrimination in language 
  3. Create models that can be applied as pre- or post-processing steps to many Natural Language Processing (NLP) systems. 

Use Cases

  1. Direct implementation in Smart Reply feature to provide ungendered response options. 
  2. Creation of a gender-balanced corpus for other NLP models to reduce bias in textual data.