Emotion Classification of Mixed-Source Hindi Language Text Using Deep Learning

In a world where Hindi ranks as the third most spoken language globally, boasting over 609 million speakers, and with a surge in user-generated content, the demand for effective emotion classification becomes more evident than ever. This linguistic landscape spans social media platforms, blogs, content reviews and numerous online interactions. Yet, despite the abundant nature of Hindi language content, a notable gap exists in emotion analysis due to the predominant focus on singular sources of textual datasets in previous studies. There have been several efforts to classify emotions in Hindi language data but they have only focused on data from single sources. Due to the limited scope of previous approaches, there is a need to study emotion classification for an aggregate of data from multiple sources. This research aims to bridge the gap by combining data from different sources and training it on various emotion classification algorithms, including traditional NLP models (such as Support Vector Machine, Random Forest, Decision Tree, K-Nearest Neighbour, Logistic Regression, Gradient Boosting and Multinomial Naive Bayes) and a fine-tuned BERT transformer. The fine-tuned BERT model ultimately achieved the highest accuracy of 83.3% in classifying emotions on the mixed-source Hindi dataset. Traditional machine learning models like SVM also demonstrated strong performance, highlighting their continued value in resource-constrained scenarios.