The School of Computer Science is pleased to present…
Conditional Query Refinement: A Type-Aware Approach with Automated Query Type Detection
PhD Dissertation Proposal by: Zahra Taherikhonakdar
Date: Tuesday, August 26, 2025
Time: 2:00 PM
Location: Essex Hall, Room 122
Web search remains the most widely used gateway for information-seeking across domains such as education, health, and e-commerce. However, user queries are often short, ambiguous, or underspecified, leading to suboptimal retrieval performance. Query refinement techniques address this challenge by modifying the initial query—through term addition, deletion, or substitution—to better capture the user’s intent. While users frequently reformulate their queries to meet evolving information needs, most neural-based query refinement methods ignore the underlying query type—such as factual, navigational, or transactional—resulting in suboptimal refinements. In this proposal, we leverage Conditional Query Refinement to improve retrieval efficacy. It shows great promise in tailoring query reformulations to better align with user intent, particularly when incorporating query type information. Our approach uses a conditional transformer architecture fine-tuned on the ORCAS-I-2M dataset, which significantly improves retrieval performance compared to type-less methods. These gains are most pronounced for navigational queries in terms of retrieval accuracy and for factual queries in generating more effective refinements. The results demonstrate that queries of different types benefit from distinct refinement strategies, highlighting the need for explicit type awareness in query refinement models.
To better address real-world applicability and scalability, we propose two key extensions. First, we will develop an AI agent leveraging large language models (LLMs) to perform automated query type detection at scale, eliminating the need for costly and inconsistent manual annotations. Second, we will explore the cross-domain generalization of type-aware query refinement by evaluating its adaptability across diverse domains.
Internal Reader: Dr. Jianguo Lu
Internal Reader: Dr. Luis Rueda
External Reader: Dr. Tanja Collet-Najem
Advisor: Dr. Hossein Fani, Dr. Ziad Kobti
