More about HKUST
Data Matching in the Era of Language Models: From Data Preparation to Data Access
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
PhD Thesis Defence
Title: "Data Matching in the Era of Language Models: From Data Preparation to
Data Access"
By
Miss Shiwen WU
Abstract:
Data matching, the task of aligning records, text fragments, or query intents,
constitutes a foundational imperative of modern data management and AI
infrastructure. While Language Models (LMs) and Large Language Models (LLMs)
have substantially advanced semantic matching capabilities, their practical
deployment is severely hindered by prohibitive annotation barriers under
low-resource regimes, task performance instability, and excessive inference
latency. To overcome these limitations, this thesis presents
resource-efficient and robust data matching frameworks tailored to backend
data preparation and frontend data access.
For backend entity resolution, we first propose CLER, an end-to-end
co-learning framework that exploits the intrinsic, complementary signals
between the blocker and the matcher. Through confidence-aware pseudo-labeling,
transitivity refinement, asymmetric selection, and dynamic sample
re-weighting, CLER mitigates label noise and breaks annotation bottlenecks
without relying on external knowledge. To further harness foundation models
cost-effectively during data preparation, we introduce SAiLER, a framework
that selectively invokes LLMs as annotators only for candidate pairs
exhibiting model disagreements or logical transitivity violations. By coupling
transitivity-guided human error correction with noise-tolerant loss weighting,
SAiLER matches fully supervised accuracy with minimal human labeling.
For frontend interactive data access, we address the recall bottleneck of the
recall-then-verify paradigm by designing TyRef, an on-demand, type-specific
query reformulation framework. TyRef diagnoses semantic discrepancies into
synonymy, conceptual hierarchies, and format variations, dynamically grounding
expansions via verified table feedback while employing adaptive triggering and
early stopping to preserve sub-second latency. Finally, we translate this
algorithmic framework into FilterPilot, an interactive demonstration system
that enables real-time inspection, latency budgeting, and human-in-the-loop
steering. Together, these contributions bridge the gap between foundation
models and production systems, delivering scalable, robust, and cost-effective
data matching across the data lifecycle.
Date: Monday, 21 September 2026
Time: 2:00pm - 4:00pm
Venue: Room 3494
Lifts 25/26
Chairman: Prof. Stuart Arthur GIETEL-BASTEN (SOSC)
Committee Members: Prof. Xiaofang ZHOU (Supervisor)
Prof. Raymond WONG
Dr. Wilfred NG
Dr. Jia LIU (MARK)
Prof. Wenjie ZHANG (UNSW)