More about HKUST
Legal Data Augmentation in LLMs for Low-resourced Domains
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
MPhil Thesis Defence
Title: "Legal Data Augmentation in LLMs for Low-resourced Domains"
By
Mr. Tsz Ho LI
Abstract:
Large language models (LLMs) have currently enhanced their reasoning
capabilities. Recent work aims to leverage reasoning skills in the legal
domain, providing legal services and developing legal frameworks using LLMs.
However, these works primarily focus on major jurisdictions with
well-annotated datasets (e.g., the US, China, and the EU), while legal
domains with less-annotated data are rarely explored. In these under-
represented regions, data scarcity limits the viability and application of
existing legal frameworks. In this thesis, we introduce CLAug, a cross-
regional legal data augmentation method, to address the data scarcity in
low-resourced legal domains. Our method uses foreign legal data, particularly
from major jurisdictions, as augmentation sources in the local legal context.
We process foreign data through a construction pipeline, so we can treat
these court cases as if they happened locally for augmentation by
transferring legal information from a foreign region to a local region. Using
the existing court cases of the European Union about privacy and AI safety,
we demonstrate effective legal knowledge augmentation on the Hong Kong
privacy law domain, under the Hong Kong Personal Data (Privacy) Ordinance
(PDPO), by in-context learning experiments using LLMs. Our results show that
LLMs can effectively learn from the augmented data and perform more accurate
legal judgments in the local context, showing the advantages of data
augmentation. We also observe two key factors that are critical to the
augmentation process: diversity from augmentation sources and locality from
original local data. These findings suggest that legal data augmentation
using well- annotated data is a promising approach to resolve the data
scarcity in low-resourced domains.
Date: Monday, 17 August 2026
Time: 2:00pm - 4:00pm
Venue: Room 3494
Lift 25/26
Chairman: Prof. Song GUO
Committee Members: Dr. Yangqiu SONG (Supervisor)
Dr. Dan XU