Twitter Sentiment Analysis in Resource Limited Language

Sentiment analysis is essential for understanding public opinion and user feedback in various languages. However, a language barrier often limits the use of existing models which are primarily pre-trained on English. Therefore, previous approaches have focused on building language-specific models for non-English languages. In this work, we investigate the efficacy of low resource language specific models (like GreekBERT) and compare their performance with RoBERTa model for predicting sentiments in Greek and English language tweets. More specifically, we explore whether Greek tweets translated to English and fed to RoBERTa model performs better than Greek tweets directly fed to GreekBERT model. We find the RoBERTa model performs well not only for the English tweets but also for the non-English tweets (Greek) translated to English. We present a detailed summary of model performance for sentiment classsification of non-English (Greek) tweets.