Skip to main content

SoBigData Articles

Aggressive Men and Sensitive Women: Measuring stereotyping bias in Large Language Models.

In the past year I had the great pleasure of visiting Universitat Pompeu Fabra, and my friend and colleague Professor Carlos (Chato) Castillo, to work on Modelling and Measuring Algorithmic Bias and Fairness. My visit had two parts, one in April of 2023 and one in February of 2024. This was the first time I visited the UPF Campus, and I was very impressed by the industrial and urban design of the buildings and the neighbourhood, as well as by the energetic and relaxed atmosphere of the campus.

 

A tall stone pillar in front of a building

Description automatically generated

 

During my visit, we had lengthy discussions with Chato around algorithmic bias and fairness, and I had the opportunity to tap into his extensive knowledge on data collection and bias measurements. We also discussed about a potential project on measuring stereotyping bias in Large Language Models, and we performed initial anecdotal measurements. We were excited to discover cases where LLMs appear to support stereotypes, such as that an aggressive person is likely to be male, while a sensitive one is likely to be female. Based on these initial observations we designed a methodology for systematically measuring bias in LLMs, which will be the basis for future experiments.

I also had the opportunity to meet with the students and faculty in the Web Science and Social Computing group at UPF. Specifically, I met with Ioannis Bilionis, Marina Almenzar, Ilaria Taddei and Abolfazi Mohammadi Seif, who described their current research at UPF. I also had very interesting discussion with Michele Starnini, and some industry collaborators of Chato. I was very motivated by the fact that the industry practitioners are interested in issues around bias and fairness, and specifically about bias in LLMs.

The visit had an amplifying effect to my research, as it kick-started a project I am very passionate about, which is to measure bias in LLMs. The discussions I had with Chato will provided a frame for pursuing this project. I also had the opportunity to expand my professional network and meet several interesting researchers.

As a side-benefit I got the opportunity to visit Barcelona again, one of the most vibrant and colourful cities of Europe. I lost myself in the streets of the old town, and I was awed again by Sagrada Familia. In my last day in Barcelona I was fortunate to experience the Lights Art Festival in the neighbourhood around the campus.

 

 

Overall, my TNA visit was a great experience both professionally and personally. It would not be possible without the SoBigData++ support, and the help of the administrative staff that made the whole process extremely smooth. I hope I have the opportunity to repeat such a TNA visit in Barcelona or elsewhere.