Evaluation of bias and robustness in search and conversational systems

Abolghasemi, A. (2026)

Doctoral Thesis

Search and conversational systems have become central to how people access information and perform tasks. With the emergence of large language models (LLMs), information systems have shifted from purely retrieval-based pipelines toward generation and retrieval-augmented generation (RAG). While these advances bring new opportunities, they also introduce challenges such as outdated knowledge, hallucinations, bias, and failures across multi-stage information systems. Ensuring that such systems are robust, unbiased, and trustworthy requires systematic evaluation across a broad range of tasks and contexts. 

In this thesis, we investigate how retrieval and generative models behave in nuanced real-world information-seeking scenarios, with a particular focus on robustness and unbiasedness, as essential aspects of building reliable and trustworthy systems.

• All authors: Abolghasemi, A.
• Supervisor: Verberne, S.; Rijke, M. de
• Co-supervisor: Azzopardi, L.
• Committee: Bonsangue, M.M.; Kraaij, W.; Hanbury, A.; Kanoulas, E.; Maistro, M.
• Qualification: Doctor (dr.)
• Awarding Institution: Leiden Institute of Advanced Computer Science (LIACS), Faculty of Science, Leiden University
• Date: 2026-03-06
• DOI: doi:10.60602/1887-4296728
• Sponsorship: Horizon 2020(H2020)
• Grant number: 860721