Evaluating Ensembled Transformers for Multilingual Code-Switched Sentiment Analysis

Sentiment analysis is essential for understanding human-authored texts, especially in multilingual communities where code-switching is common. Most existing research focuses on single-language pair sentiment analysis. We introduce a three-step approach for sentiment analysis on code-switched data: translating the code-switched data into English at word and sentence levels, training on Transformer models, and utilizing a stacking classifier to ensemble the models for sentiment classification. We establish a performance benchmark for binary and ternary sentiment classification by applying this to five datasets featuring English mixed with Spanish, Tamil, Telugu, Hindi, and Malayalam. Our method emphasizes the potential of ensembled Transformer models in this domain, paving the way for future advancements.