Sentiment Analysis of Social Media Posts using Fine-Tuned Bert Models

Mustapha Oluwatoyin & Alawode Ademola John, Volume 6 Issue 2, December 2025 Pages 34-41, Published: 2025-12-12

Abstract

As social media networks like Twitter and Facebook rapidly multiply, huge quantities of user-created content have made it possible to gain a real-time sense of the opinions held by the masses. The comprehension of this feeling is vital to many spheres, such as market research, political analysis, and crisis management. Sentiment Analysis (SA) allows deriving emotion and opinion-guided information out of the texts, but standard machine learning models have had problems coping with informality and contextual richness of social media language. This paper discusses the use of fine-tuned Bidirectional Encoder Representations from Transformers (BERT) as a sentiment classification mechanism to social media text. The BERT model was fine-tuned and tested on a multi-class sentiment task using the Sentiment140 as a large corpus of labeled tweets. The model had an accuracy of 88 percent and had a macro-averaged F1 score of 87.3 percent, which was highly above what is offered by conventional algorithms like Naive Bayes and Support Vector Machines. Findings suggest that bidirectional properties of the BERT attention process provide more contextual values of sentiments, particularly in latent and borderline phrases. It is therefore the truth that transformer-based models are effective in sentiment analysis and indicate that fine-tuned BERT is scalable and more accurate for the classification of public opinion in the social media setting.