Comparative Study of Different Learning Paradigms for Zero-Shot Sentiment Analysis of the Low-Resource African Language Oromo

Abstract

In this paper, we address zero-shot sentiment analysis for Oromo, a low-resource language spoken in East Africa, as part of SemEval-2023 Task 12 (Zero-Shot on Oromo). Leveraging large-scale language models, including BERT and its multilingual variants, we investigate four learning paradigms: zero-shot transfer, translation-based, cross-lingual, and unsupervised approaches. We conduct a comprehensive evaluation of these approaches on the SemEval-2023 benchmark and analyze their respective strengths and limitations. The results highlight the effectiveness of zero-shot transfer and translation-based methods while revealing the challenges faced by cross-lingual and unsupervised methods in preserving sentiment-specific information under zero-shot conditions. Additionally, we discuss the potential implications of our findings and outline directions for future research.