Zero-Shot Classification Reveals Potential Positive Sentiment Bias in African Languages Translations

Hrishav Sapkota, Saurav Keshari Aryal, Howard Prioleau

Keywords: Zero-Shot Classification, African Language, Sentiment Classification, Language Translations, AfriSenti-SemEval

TL;DR: Failed Zero-shot classification of African tweets translated to English reveals potential loss of sentiment from Translation

Abstract: Natural Language Processing research into African languages has been limited, with over 2000 languages still needing to be studied. We employ the AfriSenti-SemEval dataset, a recently released resource that provides annotated tweets across 13 African languages, for sentiment analysis to address this. However, given the persistent data limitations for specific languages, we translate each language to English and conduct zero-shot classification using a large BART model trained with three candidate labels: positive, neutral, and negative. Intriguingly, our findings indicate that all tweets are classified as positive. Further investigation into prediction probabilities reveals that translation technologies may exhibit a bias in translating African languages toward positive sentiments. This observation highlights the potential impact of translation tools on sentiment analysis and warrants further examination.