A Multilingual Sentiment Lexicon for Low-Resource Language Translation using Large Languages Models and Explainable AI

Addressing AI gaps in underserved regions: The research tackles a critical problem where languages spoken by millions in South Africa and the DRC—including Zulu, Sepedi, Tshiluba, and French—lack the labeled training data needed for accurate AI systems, demonstrating how responsible AI must account for linguistic diversity beyond English-dominant datasets.

High performance with transparency: The BERT model achieved 99% accuracy and 98% precision in sentiment classification while using Explainable AI (XAI) to clarify its predictions, showing that responsible AI systems can be both highly effective and interpretable—allowing users to understand why the system made specific decisions rather than treating it as a "black box."

Cultural relevance matters for accuracy: By integrating language-specific sentiment scores into the multilingual lexicon, the researchers demonstrated that sentiment analysis cannot be one-size-fits-all; culturally and linguistically tailored approaches improve both accuracy and fairness for underrepresented communities.

Practical solution for low-resource languages: The development of a comprehensive multilingual sentiment lexicon and testing corpus provides a reusable resource for future AI applications in these languages, addressing the responsible AI principle of ensuring technology benefits populations that have historically been excluded from AI development.