Context-Aware Sentiment Analysis in Setswana

Sentiment analysis traditionally faces significant challenges in low-resource languages, primarily due to the lack of robust, annotated datasets and tools tailored to specific linguistic contexts. This paper addresses these challenges within the Setswana language by leveraging its unique cultural and linguistic attributes that influence sentiment expression. We present a novel adaptation of transformer-based models, specifically PuoBERTa, bert-base-multilingual-cased, XLM-Roberta, afro-xlmr-base, multi-qa-mpnet-base-dot-v1, and setu4993/LaBSE, enhanced with contextual hints. These contextual hints are specially designed tokens embedded in the dataset to highlight key sentiment-bearing phrases, curated by native Setswana speakers. Our approach demonstrates that incorporating these contextual hints significantly enhances model performance, with PuoBERTa achieving an accuracy of 86.6%, an F1-score of 86.5%, and a Cohen's Kappa of 0.600. The findings suggest that the contextual hint methodology not only improves sentiment analysis in Setswana but also offers a scalable strategy for other under-resourced languages. This study contributes to the broader field of computational linguistics by illustrating how tailored data preprocessing combined with advanced machine learning techniques can effectively overcome the limitations posed by low-resource conditions.