[UCL WI Talk]: Re-Thinking Query Expansion in the Era of Large Language Models

Abstract

This talk will examine how we can improve query effectiveness through expansion models in the era of Large Language Models (LLMs). Firstly, we tackle the problem of poor pseudo-relevance feedback by applying an LLM-based re-ranker prior to query expansion under various conditions. We show that vast effectiveness gains can be achieved by creating a fine-grained word and entity query expansion model without additional re-ranking computation through our 'adaptive' framework. Secondly, we use LLMs directly for query expansion by introducing generative-relevance feedback (GRF). We show that this paradigm significantly improves over pseudo-relevant feedback (PRF) across sparse, dense, and learned sparse retrieval. Furthermore, we reduce the impact of generating non-relevant content through fusion and semantic grounding to the target corpora. Overall, this body of work shows how we can harness LLMs creatively to improve retrieval effectiveness significantly through query expansion.

Bio

Iain Mackie is completing his PhD at the University of Glasgow, sponsored by Bloomberg AI. His research focuses on using Large Language Models (LLMs) for query expansion under the supervision of Alan Turing fellow, Dr Jeff Dalton. Iain has several publications across information retrieval, knowledge graphs, and NLP at top academic venues (SIGIR, CIKM, AKBC, SigDIAL, etc.). He was also part of the University of Glasgow's winning Amazon Alexa Challenge team. Outside the lab, Iain is the CEO and co-founder of Malt AI, a VC-backed startup focused on using knowledge distillation for enterprise applications. Previously, he was a Quant Trader at BNY Mellon, Researcher at Edinburgh Partners, and Head of NLP at Creator Fund. Iain also currently sits on the board of Series A neural search startup Marqo.