Improving Sentiment Analysis Accuracy In Hinglish Text Using Hybrid Deep Learning Approaches

Abstract

The current study presents two hybrid deep learning models, designed and implemented to accomplish the task of sentiment analysis in Hinglish - a mixture of Hindi and English. First of them, combining Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) - pattern recognition capabilities as well more temporal dependencies than forward looking models. The second model expands on this and merges BERT, CNN with RNN (LSTM units) as well to benefit from the deep context-awareness of BERT. 5,050 Hinglish reviews pertaining to a positive one set and negative on the other aside neutral sentiments dataset was preprocessed intensively. The BERT-based model achieved 94.1% accuracy on the validation set, which outperformed CNN+ RNN model that had an accuracy of 92.97%. They generally extended the performance of best hybrid models ( tuned with considerable hyperparameters ) to mixed-language sentiment analysis.