Enhancing Monolingual Sentiment Classification: Pioneering Strategies in Tailored Language Training and Analytical Techniques

Sentiment analysis, a burgeoning field within natural language processing, has predominantly concentrated on the English language in recent years. This research conducts an in-depth qualitative review of significant research endeavors in this domain, scrutinizing the strengths and weaknesses of existing methodologies. Notably, it includes an overview of techniques adapting English resources for sentiment analysis in African languages using machine translation or transfer learning. Africa, boasting unparalleled linguistic diversity with over 2,000 languages across more than six language families, includes 75 languages with at least a million speakers each. Yet, African languages have seen limited exploration in NLP research. In our work, we introduce the AfriSenti sentiment analysis benchmark, featuring tweets in six African languages (Hausa, Igbo, Kinyarwanda, Nigerian Pidgin, Swahili, and Yoruba) spanning four language families. This focuses on task A monolingual sentiment classification datasets. Native speakers annotated these tweets, contributing to the AfriSenti-SemEval 2023 shared task 12. We observe that the most successful strategies among partici-pants involved ensemble methods and BERT-like models, proving their efficacy in sentiment analysis tasks, even for languages with scarce linguistic resources.