Sentiment Analysis for the African Language Twi: Translation-based vs. End-to-End Approaches

Abstract

This paper presents our approach to sentiment analysis for Twi, a low-resource African language. We developed two types of systems: a translation-based system and an end-to-end system. These systems leverage various popular large-scale language models (LLMs), such as BERT and its multilingual variants, and their performances were compared. Our evaluation focused on the accuracy and robustness of these systems in identifying sentiments within Twi text. We also explored the challenges associated with low-resource languages, including limited annotated datasets and the need for effective cross-lingual transfer. The results highlight the potential of end-to-end multilingual LLMs for low-resource languages while emphasizing the importance of translation quality in translation-based approaches to sentiment analysis tasks. Additionally, we provide insights into the practical implications of our findings for future research.