Sentiment Analysis for Shona

No sentiment analysis system existed for Shona yet—even though it is a Bantu language spoken by nearly 17 million people. Consequently, we collected ShonaSenti—a new corpus of 16,000 tweets in Shona covering 8 different topics. In this paper, we describe our distant supervised labelling strategies to support the annotators to categorize the collected tweets into the 5 sentiment classes of very negative, negative, neutral, positive and very positive. Moreover, we leveraged the Shona sentiment analysis corpus to develop first mono-lingual and cross-lingual sentiment analysis systems for Shona. Our best sentiment analysis systems are cross-lingual and mono-lingual Transformer-based systems which achieve accuracies of 84% on 3 sentiment classes and 70% on 5 sentiment classes.