An Opinion Predictor Using Recurrent Neural Networks

Aims: Netizens share their personal experiences, opinions at the review sites, discussion groups, blogs, forums and etc. With the rapid growth of technology, now-a-days almost everyone uses internet. Opinions are important because whenever a person needs to take a decision, helikes to hear others’ opinions. Quotes of the attitude may be generally positive or negative. We propose a system for classifying text sentiment based on Neural Networks classifier. In this paper, we focus on classifying product reviews according to the opinion and the value judgment they posses, into two polarities, positive and negative, using the multilayer neural network.
We also address opinion prediction application for the products that are being launched in future. The product features, given as input to recursive neural network are used to predict the opinions, which are expected from customers. The opinion prediction is done using recurrent neural network with the help of back propagation with time (BPTT) algorithm.
Place and Duration of Study: Department of Computer Science and Engineering, Sri Sairam College of Engineering, Anekal, Bangalore between July 2014 and December 2014.
Methodology: We experimented on 500 opinions, among them 400 were used as training set, and 100 were taken to be testing set, for each type of mobile (Nokia Lumia 720, LG G3).
Results: For each mobile type we achieved up to 85% of correct classification of opinion reviews.
Conclusion: We presented a system for determining text sentiment of product reviews by classifying them using Neural Network. The method uses feed-forward Neural Network with ten hidden layers. From the presented results, it can be seen that, a new approach is developed categorizing product reviews in 2 classes in the context of opinion mining. Experiments conducted on training sets show that with our approach we are able to extract relevant feedback from a specific domain of products. We compared our proposed opinion classification algorithm to standard algorithm BPNSO which showed the results are good between 60% to 80%.