International Journal of Advanced Research in Computer and Communication Engineering

Abstract: Kanglish is one of the common used mixed language, usually used in social media to convey messages, written using english letters that sounds in Kannada, similar to Hinglish. This research is done to analyze the sentiments of different words and sentences based on the input kanglish sentences using mBERT model which is an deep learning technique. Many sentences were collected from various media platforms to prepare a dataset labeled with different emotions. The data was divided to train, test and validation set to train and test the model. Navarasa- the nine different emotions that are potraided in classical dance with different expression can also be expressed in words. Total of 12 different emotions are being labeled and the sentiment prediction model can predict the emotion of statement. AdamW optimization, cross entropy loss and earling stopping were used to prevent overfitting. Evaluation was carried out to test the performance and its accuracy. A performance with 0.92 score was achieved with the confusion matrix that highlighted the model’s capacity to differentiate among different emotions. Gradio is used as a user interface. The results shows the potential of the transformer based model architectures for improving the sentiment analysis for the languages that are not well resourced.

Keywords: Kanglish, mBERT, AdamW optimization, cross entropy loss, early stopping, Navarasa sentiment prediction.

Downloads: | DOI: 10.17148/IJARCCE.2025.14911