Work

Public Opinion Analysis Using Sentiment Analysis: A Comparative Benchmark of Lexicon, Machine Learning, and Transformer Architectures

Year: 2026
Type: article
Abstract: Understanding public sentiment across political, civic, and consumer domains is pivotal for governance, policy formulation, and commercial strategy. Traditional public polling methodologies suffer from sampling bias, high expenditure, and non-trivial temporal latency. Microblogging and digital forums provide a continuous stream of public discourse, yet extracting structured sentiment from unorganized text poses substantial computational challenges. This paper implements a systematic 6-step analytical framework to evaluate sentiment extraction models on diverse public opinion datasets (POD-Pol24, POD-Civic, and POD-Cons). We compare four computational paradigms: lexicon-based inference (VADER), classical statistical machine learning (Linear SVM with TF-IDF), recurrent neural networks (Bi-LSTM with contextual embeddings), and fine-tuned self-attention Transformers (RoBERTa-base). Experimental metrics demonstrate that while lexicon methods yield rapid baseline inference (F1 = 0.681), fine-tuned Transformer models achieve superior contextual accuracy (F1 = 0.923), effectively resolving polysemy and complex sentence structures. We further examine domain-specific constraints including sarcasm, code-switching, and computational scalability.
Institution: Srinivas University
Language: English
Open Access status: green
OpenAlex ID: W7234214979
FWCI:
Cites:
Cited by:
Related to: 10