Voting Ensemble Learning for Multilingual Sentiment Analysis: A Comparative Study

Abstract

Multilingual sentiment analysis (MSA) faces many challenges due to language differences, cultural considerations, and data sparsity across languages. Although ensemble learning shows promise for improving robustness and accuracy by combining classifiers, there are relatively few systematic comparisons of different ensemble approaches, including Max Voting and bagging, boosting, and stacking using modern transformer models for MSA. This study fills that gap with an empirical examination in its contribution here. We assessed the Max Voting ensembles based on diverse Transformer models (e.g., LaBSE, DistilBERT, and XLM-R variants) that were fine-tuned on a multilingual dataset, specifically Twitter. We took our time to compare Max Voting against bagging, boosting (XGBoost), and stacking ensembles to analyze the scenarios where Max Voting performed best, most notably when the base models had solid and stable predictions. The study shows that Max Voting gives a competitive and consistent performance that rarely exceeds optimal performance with only two to four models, thus ensuring efficiency. While complexity methods, like boosting and stacking, can achieve higher performance in a few cases, Max Voting serves as a highly effective baseline that also simplifies the analysis. Additionally, this work looks to identify language-performance considerations at a baseline level, and provides readers with practical, data-driven guidance to assess ensemble development guided by levels of accuracy, computational constraints, and linguistic considerations. Ultimately, the findings can assist readers in real-world applications incorporated into multilingual domains.