Machine Learning-Based Binary Sentiment Classification of Movie Reviews in Hindi (Devanagari Script)

Lately, there has been a remarkable surge in online movie reviews in Hindi with the advent of the UTF-8 standard. Movie reviews are an excellent source of sentiments; therefore, Hindi movie review classification is one of the exciting and demanding tasks of NLP, as it helps the viewers decide whether a film/movie is worth watching. Much work in movie reviews sentiment classification has been done mainly for resource-affluent languages. Still, preliminary work is being done in Hindi due to its complex nature and scarce resources like adequate-labeled datasets. This paper aims to develop a machine learning-based solution for performing binary sentiment classification on movie reviews in Hindi (Devanagari Script). To this end, a primary binary polarity dataset, namely, Movie Reviews in Hindi (MRH) consisting of 5K reviews, is made. Apart from MRH, the Hindi IIT-P movie and product review datasets are also deployed in this work. Firstly, all three datasets are prepared for further processing using the preprocessing steps, and the features used are unigram, bigram, and trigram, along with TF-IDF. Second, various state-of-the-art classifiers are applied to all three datasets. Further, we proposed and used a stacked model of classifiers for performing binary sentiment classification on Hindi reviews. Experimental results on all three datasets prove that the proposed stacking ensemble based on the employed features compared favorably to all the baseline classifiers applied and achieved reasonably high performance. Therefore, it indicates the efficacy of the proposed stacked model for sentence level movie reviews sentiment classification in a resource-scarce scenario.