Fairness of AI Systems
Supervisor
Suitable for
Abstract
Having fair AI systems is critical for the deployment of these systems in society. Measuring the fairness of a system is
currently achieved, for example, by using diagnostic datasets [1]. However, current diagnostic datasets may still be unfair
[2]. The goal of this project is to investigate the potential sources of unfairness in existing diagnostic datasets, and to
propose solutions for fixing these datasets or to gather new diagnostic datasets grounded in linguistic theory that would
not suffer from unfairness. For example, for assessing gender bias in NLP models, current diagnostic datasets (e.g., GAP [1],
WinoGender [3], WinoBias [4]) are either synthetic or have unbalanced properties that do not allow for correct bias measuring
[2]. One could gather a new real-world dataset for assessing gender bias in a similar way GAP was gathered [1] (by selecting
paragraphs from Wikipedia that adhere to certain rules), but, additionally, restricting to instances where it is possible
to obtain a counterpart with swapped genders.
The students will also be able to propose their own projects in this
area. The project will also be co-supervised by Oana-Maria Camburu (postdoctoral researcher) and Vid Kocijan (final year DPhil
student), who have strong background and contributions in this area.
Strong coding skills (preferably in deep learning platforms, such as Pytorch or Tensorflow), deep learning knowledge.
References:
[1] https://www.aclweb.org/anthology/Q18-1042.pdf
[2] https://arxiv.org/abs/2011.01837
[3]
https://arxiv.org/abs/1804.09301
[4] https://arxiv.org/abs/1804.06876