Understanding and Mitigating Gender Bias in Information Retrieval Systems

Abstract:

Recent studies indicate that information retrieval systems can display stereotypical gender biases, potentially leading to discrimination against minority groups and influencing users’ decisions and judgments. In this talk, I will discuss systematic studies that reveal the presence of stereotypical gender biases in Information Retrieval (IR) systems. I will also categorize existing research on gender biases in IR systems, focusing on (1) relevance judgment datasets, (2) the structure of retrieval methods, and (3) the representations learned for queries and documents. Drawing on these categories, I’ll present a range of methods that my collaborators and I have developed to measure, control, or reduce stereotypical biases in IR systems. I’ll also introduce key datasets and collections commonly used to study gender biases in IR systems, along with evaluation metrics for assessing model bias and utility and discuss de-biasing techniques to help mitigate gender biases within these models.

Biography:

Shirin Seyedsalehi is a fourth-year Ph.D. student in Computer Engineering at Toronto Metropolitan University, supervised by Dr. Ebrahim Bagheri and Dr. Morteza Zihayat. She earned her B.Sc. and M.Sc. in Electrical Engineering from Amirkabir University of Technology (Tehran Polytechnic). Shirin’s research centers on mitigating stereotypical gender biases in information retrieval systems, with a focus on neural rankers. She has published in top-tier venues such as Machine Learning Journal, SIGIR, ECIR, CIKM, and EDBT and has served as a program committee member for conferences including SIGIR 2023, CIKM 2024, SIGIR 2024, ECIR 2024, and SIGIR-AP 2024. Her forthcoming book on Gender Bias in Information Retrieval will be published by FnTIR.