Emotion recognition in Hindi text using multilingual BERT transformer

Emotions are a vital and fundamental part of our existence. Whatever we do, say, or do not say somehow reflects our feelings, however not immediately. To comprehend human's most fundamental behaviour, we must examine these feelings using emotional data. According to the extensive literature review, categorising speech text into multiple classes is now undergoing extensive investigation. The application of this research is very limited in local and regional languages such as Hindi. This study focuses on text emotion analysis, specifically for the Hindi language. In our study, BHAAV Dataset is used, which consists of 20,304 sentences, where every other sentence has been manually annotated into one of the five emotion categories (Anger, Suspense, Joy, Sad, Neutral). Comparison of multiple machine learning and deep learning techniques with word embedding is used to demonstrate accuracy. And then, the trained model is used to predict the emotions of Hindi text. The best performance were observed in case of mBERT model with loss- 0.1689 ,balanced_accuracy- 93.88%, recall- 93.44%, auc- 99.55% and precision- 94.39 % on training data, while loss- 0.3073, balanced_accuracy- 91.84%, recall- 91.74%, auc- 98.46% and precision- 92.01% on testing data.