SFI researchers claim success in reducing gender bias in AI models

9 Jun 2022

Image: © knssr/Stock.adobe.com

Researchers at the Adapt centre say their new research approach can work with multiple languages with minimal changes, decreasing the cost and time spent on mitigating bias.

Researchers at the Science Foundation Ireland Adapt centre claim they can reduce gender bias in natural language AI more efficiently than current methods.

Machine learning algorithms are built on the training data that they receive, but there can be human biases within language data. These biases can alter how a natural language model functions, resulting in the same mistakes or assumptions being made constantly.

Mitigating bias in natural language processing requires large amounts of gender-balanced training data, which increases the cost and time to deliver a model.

However, a team from the Adapt research centre for AI-driven digital content technology has leveraged pre-trained deep-learning language models to make this easier.

They claim this new research approach accelerates effectiveness, and could therefore make the development of AI language models more affordable and less time-consuming.

“From finding a romantic partner to getting that dream job, artificial intelligence plays a bigger role in helping shape our lives than ever before,” said Adapt research engineer Nishtha Jain, who led the research.

“This is the reason we as researchers need to ensure technology is more inclusive from an ethical and socio-political standpoint. Our research is a step in the direction of making AI technology more inclusive regardless of one’s gender.”

Adapt’s new approach to developing this technology is designed to work on multiple languages with minimal changes in the form of heuristics. To test this claim, the research was trialled on Spanish, which has a large amount of available data, and Serbian, which is a low-resource language. The team said they achieved positive results.

The research was done in collaboration with Microsoft and Imperial College London and will be presented at a European Language Resource Association conference later this month.

Biases in AI continue to be a concern when developing new software. Last month, Google said a preliminary analysis of its text-to-image system Imagen encoded a range of “social and cultural biases” when generating images of activities, events and objects.

10 things you need to know direct to your inbox every weekday. Sign up for the Daily Brief, Silicon Republic’s digest of essential sci-tech news.

Leigh Mc Gowran is a journalist with Silicon Republic