From Comments to Insights: Comparing Local and Online LLM Models in Social Media Sentiment Evaluation

Riza Akhitova, Novikov David Gennadievych, Akbergen Dinara Sadyrbekkyzy, David Novikov

2025

Closed access Metadata (CSL-JSON)
DOI 10.1109/sist61657.2025.11139353
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Social media platforms such as YouTube, TikTok, and Reddit generate vast amounts of user-generated content, making sentiment analysis essential for understanding public opinion. Traditional sentiment analysis methods often struggle with sarcasm, informal language, and evolving trends, leading to the rise of Large Language Models (LLMs) as more effective alternatives. However, the choice between cloud-based LLMs (e.g., ChatGPT-4o, Gemini 2.0 Flash) and local LLMs (e.g., Meta LLaMA 3.2, DeepSeek-R1 7b) remains an open question, with each approach offering distinct advantages and limitations.This study compares the performance of online and local LLMs in sentiment analysis, focusing on discussions about AI-generated art and cancel culture. Through structured sentiment classification and comparative evaluation, we assess each model’s accuracy, efficiency, and interpretative depth. Our findings indicate that cloud-based models excel in contextual comprehension, while local models provide enhanced privacy and structured data processing. The results highlight the trade-offs between computational power, data security, and adaptability, reinforcing the potential of hybrid approaches that integrate both cloud and local LLM capabilities for more effective sentiment analysis.The goal of this research is to provide a clear evaluation of different LLMs in sentiment analysis, offering practical insights into their strengths and weaknesses. By comparing models from multiple developers, we aim to determine the most effective deployment strategies for real-world applications, ensuring accurate and scalable sentiment evaluation in diverse online discussions.