Date Approved

9-8-2026

Embargo Period

9-8-2026

Document Type

Thesis

Degree Name

M.S. Computer Science

Department

Computer Science

College

College of Science & Mathematics

Advisor

Vasil Hnatyshin, Ph.D.

Committee Member 1

Umashanger Thayasivam, Ph.D.

Committee Member 2

Serhiy Hnatyshyn, Ph.D.

Disciplines

Computer Sciences | Physical Sciences and Mathematics

Abstract

Authors of scientific papers rely heavily on acronyms and often use them without defining them, making the literature harder to read and index. This thesis develops and evaluates a hybrid rule-based and large language model (LLM) framework that extracts acronym–definition pairs from scientific PDF documents. It extends an earlier Rowan University system that combined a regular-expression parser with a single LLM on 200 papers. That system showed that neither the parser nor the LLM alone is sufficient for accurate extraction of the pairs. The framework is a fully automated pipeline from PDF input to scored results. It compares four LLM providers through their lightweight, low-cost tier models (GPT-4o mini, Gemini 2.0 Flash, Claude Haiku 4.5, and Kimi K2.5) under identical prompts, so differences reflect model capability rather than prompt design. Each LLM provider is evaluated on three input types (the raw PDF, unprocessed extracted text, and preprocessed text) and on seven method configurations spanning the parser alone, each LLM alone, the raw PDF path, and the parser refined by an LLM. Evaluation against a manually annotated ground truth of about 200 papers across four scientific domains reports precision, recall, and F1 score with confidence intervals. This evaluation confirms that preprocessing substantially improves precision, that the LLMs balance precision and recall differently, and that refining the parser output with an LLM trades precision for recall. An additional comparison of around 800 unannotated documents confirms these findings. A hybrid design with careful preprocessing yields accurate and scalable extraction of acronym–definition pairs, and the production extraction of about 10,000 documents remains to be run.

Download

Share

COinS