More about HKUST
Semantic Matching over Structured Data: From Data Preparation to Data Access in the Era of Language Models
PhD Thesis Proposal Defence
Title: "Semantic Matching over Structured Data: From Data Preparation to Data
Access in the Era of Language Models"
by
Miss Shiwen WU
Abstract:
Managing and querying structured data requires resolving semantic
discrepancies across different stages of the data lifecycle. These semantic
gaps manifest as variations in data records during backend data preparation,
and as mismatches between user intents and database values during frontend
data access. While language models (LMs) and Large Language Models (LLMs)
offer strong semantic comprehension, deploying them in real-world database
systems is often bottlenecked by high annotation costs, computational
overhead, and runtime latency. This thesis addresses these challenges by
developing resource-efficient semantic matching frameworks tailored to the
unique constraints of both data preparation and data access.
In the data preparation stage, we focus on a typical data cleaning task,
i.e., entity resolution (ER), and address the high cost of manual data
labeling. We first present an end-to-end iterative co-learning framework.
This approach utilizes smaller pre-trained LMs to jointly train the blocker
and matcher components through iteratively updated pseudo-labels,
incorporating targeted optimization techniques to filter out label noise. To
further leverage the advanced capabilities of LLMs without incurring
prohibitive API costs, we then introduce SAiLER. This framework minimizes the
need for human annotation by selectively employing LLMs to label only
uncertain data pairs or instances that violate data transitivity, using a
transitivity-guided correction mechanism to ensure accuracy while controlling
the labeling budget.
Transitioning to the data access stage, the focus shifts from offline data
cleaning to real-time query processing, where user predicates fail to match
exact database text columns due to semantic gaps. We propose TyRef, a
type-specific query reformulation framework within the Recall-Then-Verify
paradigm. TyRef categorizes semantic gaps into synonyms, hierarchies, and
format inconsistencies, applying specialized reformulation strategies. To
meet low-latency requirements, TyRef uses an on-demand triggering mechanism
that activates LLMs only when substantial semantic gaps are detected,
employing adaptive selection and early stopping to prune redundant
computation.
Together, these frameworks provide scalable and practical solutions that
systematically balance model accuracy against resource limitations,
facilitating data integration, cleaning, and querying database systems in the
era of language models.
Date: Wednesday, 29 July 2026
Time: 2:00pm - 4:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Prof. Xiaofang Zhou (Supervisor)
Prof. Raymond Wong (Chairperson)
Dr. Wilfred Ng