Date of Graduation

7-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Engineering (PhD)

Degree Level

Graduate

Department

Computer Science & Computer Engineering

Advisor/Mentor

Gauch, Susan

Committee Member

Adams, Douglas

Second Committee Member

Gauch, John

Third Committee Member

Huang, Miaoqing

Fourth Committee Member

Guach, Susan

Keywords

Modern Natural Language Processing (NLP), Information Retrieval (IR), Sequence Graph Network (SGN), Sequence Graph Attention (SGA)

Abstract

Modern Natural Language Processing (NLP) and Information Retrieval (IR) systems have achieved remarkable success in semantic understanding; however, they continue to struggle with complex logical reasoning, structural interactions, and dynamic aggregation. This dissertation presents a comprehensive framework to enhance the structural, logical, and interactive capabilities of language models and retrieval systems across four progressive studies. First, we address the challenge of modeling dynamic conversational logic in online debates. We introduce a Sequence Graph Network (SGN) that captures the temporal and interactive exchange of ideas—such as counterarguments and reinforcements—by updating node features sequentially through a novel Sequence Graph Attention (SGA) layer. Second, we transition to the logical limitations of dense retrieval models, which often fail to interpret Boolean logic and implicit set operations correctly, treating operators like “and" and “not" as synonyms. We propose SetBERT, a fine-tuned BERT model utilizing an innovative inversed-contrastive loss to significantly improve document retrieval for queries involving intersections, unions, and differences. Third, to embed this logical awareness deeper into the transformer architecture, we introduce Boolean-aware attention (BoolAttn), a plug-and-play mechanism that dynamically adjusts token-level interactions using specialized attention experts for specific Boolean operators. Finally, recognizing that robust retrieval requires fusing multiple heterogeneous systems, we propose LLM-IRA, a quantum-inspired interactive ranking aggregation framework. LLM-IRA replaces expensive listwise prompting with an efficient interactive LLM loop that samples high-disagreement document pairs, learning optimal fusion weights with a fraction of the computational cost. Together, these four contributions bridge the gap between semantic matching and symbolic reasoning, offering scalable, logically robust, and interactive solutions for next-generation NLP and IR pipelines.

Download

Share

COinS