Comparative Analysis of Transformer Models for Sentiment Analysis in Low-Resource Languages

Abstract

The analysis of sentiments expressed on social media platforms is a crucial tool for understanding user opinions and preferences. The large amount of the texts found on social media are mostly in different languages. However, the accuracy of sentiment analysis in these systems faces different challenges in multilingual low-resource settings. Recent advancements in deep learning transformer models have demonstrated superior performance compared to traditional machine learning techniques. The majority of preceding works are predominantly constructed on the foundation of monolingual languages. This study presents a comparative analysis that assesses the effectiveness of transformer models, for multilingual lowresource languages sentiment analysis. The study aims to improve the accuracy of the existing baseline performance in analyzing tweets written in 12 low-resource African languages. Four widely used start-of-the-art transformer models were employed. The experiment was carried out using standard datasets of tweets. The study showcases AfriBERTa as a robust performer, exhibiting superior sentiment analysis capabilities across diverse linguistic contexts. It outperformed the established benchmarks in both SemEval-2023 Task 12 and AfriSenti baseline. Our framework achieves remarkable results with an F1-score of 81% and an accuracy rate of 80.9%. This study provides validation of the framework's robustness in the domain of sentiment analysis across a low-resource linguistics context. our research not only contributes a comprehensive sentiment analysis framework for low-resource African languages but also charts a roadmap for future enhancements. Emphasize the ongoing pursuit of adaptability and robustness in sentiment analysis models for diverse linguistic landscapes.