More about HKUST
Constraint Modeling in Natural Language Processing in the Era of Deep Learning
PhD Thesis Proposal Defence
Title: "Constraint Modeling in Natural Language Processing in the Era of Deep
Learning"
by
Mr. Zizheng LIN
Abstract:
The remarkable success of deep learning in Natural Language Processing (NLP)
stems from neural models' ability to learn complex patterns from vast amounts
of data. However, this data-driven paradigm often struggles to incorporate
prior knowledge—ranging from linguistic rules and domain expertise to
commonsense facts and logical structures—a limitation that becomes
particularly severe in data-scarce scenarios or tasks requiring rigorous
adherence to structural and semantic dependencies. This thesis investigates
constraint modeling, a paradigm that addresses these limitations by
integrating prior knowledge into models through explicit constraints.
The central idea of this dissertation is that constraint modeling, when
strategically applied at different stages of the NLP pipeline, provides a
versatile and powerful framework for enhancing models' performance,
robustness, and generalizability. We organize our study around three distinct
paradigms: (1) imposing constraints on learning, where constraints are
integrated directly into the training objective to regularize the parameter
space; (2) imposing constraints on inference, where constraints serve as a
global reasoning layer to ensure outputs adhere to logical or domain-specific
rules; and (3) imposing constraints on prompts, where instructions specifying
constraints are inserted to explicitly guide the generation processes of
Large Language Models (LLMs).
We substantiate this taxonomy through three original contributions. First, we
present HMTGIN to address heterogeneous multi-task learning in community
question answering. HMTGIN imposes explicit cross-task constraints derived
from prior knowledge directly into the joint learning objective,
demonstrating significant performance improvements across diverse tasks.
Second, we propose GCPEAC, an approach that leverages global constraints with
prompting for zero-shot event argument classification. GCPEAC adopts
constraints that encode cross-task, cross-argument, and cross-event relations
to regularize initial predictions, achieving state-of-the-art zero-shot
performance without any training.
Third, we develop CCoToM for enhancing theory-of-mind reasoning in LLMs.
CCoToM adaptively inserts natural language instructions specifying
definitional and dependency constraints into prompts at each reasoning step,
enabling more robust and accurate theory-of-mind reasoning in LLMs.
Date: Wednesday, 16 April 2026
Time: 2:00pm - 3:30pm
Venue: Room 2132C
Lift 22
Committee Members: Dr. Yangqiu Song (Supervisor)
Dr. Wei Wang (Chairperson)
Dr. Binhang Yuan