A Novel Framework For Sentiment Analysis Using Deep Learning

Large amount of un-structured data is present online in the form of opinions and reviews. The most important task of NLP is to Extract useful information from unstructured data by first converting into structured form. Many customers write down reviews online but do not give rating to them. The main concern of this paper is to perform sentiment analysis by predicting two main types of polarities from reviews available online i-e positive and negative. Neural networks models fail to capture the contextual meaning of words and also fails to save long sequences of words and thus results in reducing performance. To overcome this issue a novel Hybrid model (RNN-LSTM-BiLSTM-CNN) using majority voting, word2vec and pre-trained Glove embedding (100d) is proposed to predict sentiment polarity against each review. Loss function used is Binary cross entropy. The proposed model is tested on different state-of-the-art datasets like SST-1, SST-2 and MR Movie review dataset. Results proved that our proposed model results in improved accuracy.