Sentiment classification in Hindi text using hybrid deep learning method

Sentiment analysis (SA) gives the tool to the researcher to evaluate the sentiments of different users in multiple languages in online discourse. In the last few years, non-English content, such as Hindi language based content, has increased rapidly on the web. However, there is a scarcity of research in Hindi language as confirmed by the literature review. In this work, sentimental analysis task on 5 classes in Hindi text is examined. Due to the absence of a standard 5-class dataset in Hindi, we first curate a suitable dataset consisting of over 100 of Amazon’s best-selling books. Next we examine popular word embeddings for text representation, followed by models such as LSTM and BiLSTM, and evaluate their performance using standard metrics. It is observed in the experiments that using the fasttext word embedding and CNN-BiLSTM model, the accuracy, precision, recall, and F1-measure obtained are 77%, 76%, 76% and 76% respectively for 5-class SA in Hindi. We also evaluate the results on standard 3-class classification task to establish robustness of the approach. In both the cases, the results outperform the recent state-of-art results.