A Novel Framework for Sentiment Analysis using Deep Learning

Abstract

With the advancement of Social media, data present online is also growing progressively. Internet is full of people opinions but mostly in unstructured form. This type of data is useful for different people who are interested to know about specific product accurately. To make this unstructured data useful for different purposes like improving business, customer satisfaction; different techniques of Natural language processing are used. Extracting useful data by converting the unstructured data to structured form is an important step in NLP. Sentiment analysis is used to analyze people sentiments available online in the form of blogs, reviews, comments etc. One of the main tasks of NLP is sentiment analysis. The main problem arises because of text-only reviews where people don’t provide the rating but write down a whole statement for a product, organization or a restaurant. And it became difficult for business analysts to know about their customers reviews and thus it become difficult for them to improve their products according to customers demand. Neural networks language models fail to deal with long sequence of words and also cannot deal with contextual information. Different machine learning models are used for prediction in literature. Recently new methodologies are introduced using deep neural networks like Recurrent neural networks (RNN). But RNN does not save the previous information, to overcome this issue Long-Short-Term-Memory (LSTM) are introduced. Other improved models include BiDirectional LSTM (BiLSTM), Gated Recurrent neural network (1) and Convoluted Neural Network (CNN). To improve the prediction accuracy and to generate more reliable results a novel hybrid model is introduced having five different models including RNN, LSTM, BiLSTM, GRU and CNN and Glove word Embedding. The use of word embedding techniques is also an efficient step in sentiment analysis task. The Provided rating is converted to positive and negative Sentiment firstly in order to predict the Sentiment related to a given review. The proposed model (RNN-LSTMBiLSTM-GRU-CNN) is tested on different state of the art datasets including SST-1, SST-2, Movie Review and IMDB movie review dataset. The Results shows an improved accuracy using proposed model.