Advanced Query Representation and Feedback Methods for Neural Information Retrieval

The School of EECS is hosting the following PhD Thesis Review Seminar:

 

Advanced Query Representation and Feedback Methods for Neural Information Retrieval

 

Speaker: Mr. Hang Li

Host: Prof. Guido Zuccon

 

 

Abstract:

The emergence of neural information retrieval (IR) systems has fundamentally reshaped how relevance is computed, by introducing dense representations and large language models (LLMs) into the core of search architectures. However, the longstanding and widely studied technique of pseudo-relevance feedback (PRF), originally designed for sparse, bag-of-words systems, remains underexplored in the context of modern neural retrievers. This thesis investigates advanced query representation and feedback mechanisms tailored for transformer-based rerankers, dense retrievers, and LLM-driven retrieval systems. The overarching objective is to enhance retrieval effectiveness while maintaining robustness and computational efficiency, particularly in zero-shot and resource-constrained environments.

 

We first explore PRF strategies for transformer-based and dense retrieval models. We begin by re-examining classical text-based PRF under neural rerankers, highlighting the impact of input-length constraints, score aggregation strategies, and susceptibility to query drift. Subsequently, we investigate vector-based PRF approaches that leverage embedding fusion, including Rocchio-style embedding fusion method, which exhibit consistent improvements across multiple dense retrievers. We also investigate a dense-sparse interpolation framework, providing practical guidance on when and how hybrid feedback signals should be integrated.

 

We then address the limitations of conventional PRF methods by introducing lightweight, adaptive alternatives. We analyze how the quality of initial retrieval signals influences downstream feedback effectiveness, revealing varying degrees of robustness among methods under degraded conditions. The aforementioned works lead to the proposal of TPRF, a transformer-based PRF approach designed for efficiency-critical environments. TPRF autonomously optimizes Rocchio-style feedback vector weight parameters and demonstrates a favorable balance between effectiveness and inference latency. We further contribute a modular PRF framework implemented within Pyserini, enabling reproducible experimentation and extensible integration with state-of-the-art dense retrieval pipelines.

 

With the emerging of generative large language models, we also explore LLM-based retrievers, and investigate in advancing PRF into this new space. We propose LLM-VPRF, a generalization of vector-based PRF tailored to decoder-style LLMs, and PromptPRF, which integrates LLM-generated passage features for dense query refinement in a zero-shot setting. LLM-VPRF shows that VPRF method still holds even for large generative models, the PromptPRF methods demonstrate that small dense retrievers, when augmented with strategically extracted LLM features, can achieve competitive effectiveness of much larger models, improving cost-efficiency without compromising retrieval quality. A comprehensive analysis across model sizes, prompt strategies, and feedback types reveals novel insights into the interplay between representation depth and feedback informativeness.

 

Lastly, we identify recurring failure modes such as query ambiguity, drift, and feedback noise. These insights inform a principled understanding of when feedback is beneficial and when it may degrade performance. We conclude by outlining directions for future work, including adaptive PRF depth selection, integration with multi-modal retrieval, and feedback-aware pre-training.

 

In summary, this thesis advances the theoretical and practical understanding of feedback-driven query refinement in neural IR, providing a coherent framework that spans traditional, dense, and LLM-based retrieval paradigms. The contributions offer both methodological innovations and empirical evidence for building robust, scalable, and effective retrieval systems in the era of neural search.

 

Biography:

Mr. Hang Li is a PhD candidate from the School of EECS under the supervision of Prof. Guido Zuccon and A/P Bevan Koopman. He received his Bachelor of Science in Computer Science degree from the University of Minnesota Twin-Cities, United States. His research interests are Relevance Feedback, Query Representation/Reformulation, and Neural Information Retrieval.