ENHANCING AUTOMATED PUBLIC OPINION ANALYSIS: A COMPARATIVE STUDY OF VADER AND LARGE LANGUAGE MODEL METHODS

Abstract

Traditional sentiment analysis methods such as VADER have long been employed togauge public opinion through rule-based and lexicon-driven approaches. However, the advent of large language models (LLMs) has introduced novel methods capable of capturing deeper semantic nuances and handling heterogeneous data at scale. In this paper, a comparative analysis of VADER and LLM-based sentiment analysis methodologies is done within the context of public policy research. Using nuclear fusion as a case study, the study evaluates how each approach processes and interprets sentiment in news and social media texts. The VADER approach, characterized by its reliance on curated sentiment lexicons and syntactic rules, excels in computational efficiency and straightforward implementation. In contrast, the LLM-based framework leverages parameter-efficient fine-tuning techniques to adapt pre-trained models for nuanced sentiment extraction, thereby capturing complex opinion dynamics and contextual subtleties that traditional methods often overlook. Through rigorous experimentation, the accuracy, scalability, and processing speed of these methodologies were compared, outlining their respective advantages and limitations. Our results demonstrate that while VADER provides rapid sentiment scoring suitable for high-volume datasets, the LLM-based approach delivers richer, more robust insights that are critical for dynamic public policy analysis. This thesis further discusses the trade-offs between computational overhead and analytical depth, emphasizing that the choice of technique should align with the specific requirements of policy research. Overall, this work contributes to a better understanding of how modern deep learning techniques can enhance automated sentiment analysis and improve the timeliness and accuracy of public opinion monitoring in contemporary policy contexts.