BRIDGING THE DIGITAL DIVIDE: A COMPREHENSIVE REVIEW OF SENTIMENT ANALYSIS FOR LOW-RESOURCE LANGUAGES

Abstract

Digital content has never been as crucial in comprehending the opinion of the masses as it is now and sentiment analysis is a major part of it. Nevertheless, there is a pronounced exception: and as well as popular languages such as English experience the privileges of sophisticated Natural Language Processing (NLP), most of the low-resource languages do not. The recent developments of sentiment analysis into these underrepresented languages are stressed in this review paper, and the most relevant ones that have been covered include African, South Asian, and other marginalized languages families. We critically analyze how previous techniques of machine learning have been replaced by the current ones which are motivated from pre-trained language models (PLMs), which are based on the transformer architecture. We talk about a number of approaches that can be used to address the issue of sparse data, including the use of cross-lingual transfer learning, multilingual adaptive fine-tuning, and creative data augmentation. An important part of the review is devoted to the topic of the increasing popularity of Explainable AI (XAI) and how the tools such as LIME and SHAP may be used to establish trust and openness in such systems. We also explore the language distinctiveness these languages have to deal with, like code-switching, dialects and complicated structure of words. We paint a full picture of the field by looking through the most important benchmark datasets available such as SAfriSenti, AfriSenti and NajjaSenti and discussing sophisticated models such as Afro-XLMR, AfriBERTa and SERENGETI. After this we finally conclude by listing the current issues and proposing future research avenues such as more efficient fine-tuning methods, bigger multilingual data sets and more ethical and human-oriented AI systems.