Deep learning techniques for sentiment analysis in code-switched Hausa-English tweets

Social media serve as a crucial platform for expressing opinions and perspectives. Its texts often characterised by code-switching or mixed languages in multilingual setting. This results in a diverse and complex linguistic context, which can negatively affect the accuracy of sentiment analysis for low-resource languages such as Hausa. Prior research has predominantly concentrated on sentiment analysis within single-language data rather than code-switched data. This paper proposes an efficient hyperparameter tuning framework and a novel stemming algorithm for the Hausa language. The framework leverages word embeddings to determine the polarity scores of code-mixed tweets and enhances the accuracy of sentiment analysis models in low-resource language. The extensive experiments demonstrate the framework's efficiency and reveal a superior performance of transformer models over conventional deep learning models. The framework achieves a balance between accuracy and computational efficiency, making it suitable for deployment in practical applications. Compared to state-of-the-art transformer models, our framework significantly reduces computational costs while maintaining competitive performance. Notably, the AfriBERTa model achieves outstanding results, with an F1-score of 0.92 and an accuracy of 0.919, surpassing current baseline standards. These findings have broad implications for social media monitoring, customer feedback analysis, and public sentiment tracking, enabling more inclusive and accessible NLP tools for underrepresented linguistic communities.