Semi-Supervised Mix-Hindi Sentiment Analysis using Neural Network

Most of the people in the world now a day would like to share and express their feelings, views, experiences, suggestions and opinions on the web. These opinions are processed by sentiment analysis task and find their polarity.In this paper, we use input text file in Devanagari script stored in UTF-8 encoding scheme. We propose 3 approaches for doing sentiment analysis for Hindi multidomain review. In approach 1, classification of data is done using NN Prediction by using pre-classified words. In approach 2, classification of data is done using IIT-Bombay Hindi SentiWordNet (HSWN). In approach 3, classification of data is done using NN prediction using pre-classified sentences as labeled data. Finally, we report accuracies in every approach. We have different domain (Health, Business, Current affairs, Tourism, Movie, Technology and Product) review dataset manually and randomly collected by us. They contain Mix-Hindi words like (brave), (careful), (mineral), etc., for which we have created dictionary to deal with them. We achieve overall accuracy of 52% in approach 1, 71.5% accuracy in approach 2 and 70.27% accuracy in approach 3.