Document Type

Dissertation - Open Access

Award Date

2026

Degree Name

Doctor of Philosophy (PhD)

Department / School

Electrical Engineering and Computer Science

First Advisor

Sungyong Jung

Abstract

Large language models (LLMs) perform well on general tasks, yet their deployment in specialized, high-stakes domains such as public health remains unreliable. Models hallucinate facts, combine multi-step evidence incorrectly, and retrieve context that corrupts downstream reasoning. Three root causes underlie these failures: training data that underrepresents rare but critical domain themes, reasoning that is opaque and unverifiable, and evidence use that cannot be re-checked when the evidence changes. This dissertation addresses all three. First, we introduce HDBSCAN_SSA, a coarse-to-fine topic-modeling pipeline that couples density-based clustering with a novel Semantic Seed Aggregation algorithm. Using samples drawn from a larger corpus of approximately 469,000 COVID-19 tweets, it discovers more coherent topics than LDA, k-means, and BERTopic, and topic-guided sampling under fixed labeling budgets of 10%, 20%, and 40% consistently improves downstream sentiment-classification accuracy and macro-F1 over random and BERTopic-based sampling. Second, we propose a two-stage logic-graph reasoning interface in which an LLM extracts a structured logic graph from a claim and its evidence table, and a graph attention network classifies the graph. On PubHealthTab, a public-health table-based fact-checking dataset of 1,942 claim-table pairs, the method reaches 88.1% accuracy and 0.853 macro-F1, outperforming chain-of-thought, zero-shot chain-of-thought, and self-ask baselines, with explanation quality (ROUGE-1 0.342, ROUGE-L 0.235) exceeding direct LLM outputs. Third, we propose a grounded verification framework that makes structured reasoning re-checkable against current evidence. A multi-agent pipeline binds extracted operations to exact table cells, and a symbolic layer checks whether the bound values still support the stated reasoning result, overriding the verdict when they do not. The framework achieves 95.6% clean accuracy, 9.8 percentage points above the strongest prompting baseline. Under evidence-side attacks it corrects 95.2% of 209 cell-swap and 91.1% of 538 numerical perturbations, and it resolves 74.9% of 538 claim-side negations through re-extraction. Together, these contributions form a data-to-reasoning-to-verification pipeline for domain AI whose outputs are grounded, inspectable, and robust.

Publisher

South Dakota State University

Download

Share

COinS
 

Rights Statement

In Copyright