Sentiment Analysis on Movie Reviews using Recursive and Recurrent Neural Network Architectures

In this project, we introduce a method to tackle the problem of sentiment classification of large movie reviews, each consisting of multiple sentences. We leverage a new architecture that encapsulates previous work applied to sentence-level sentiment classification, using Recursive Neural Networks and use it in unison with a Recurrent Neural Network (where each time step is a sentence in the review). In this report, we describe our results with this architecture as well as four other intermediate approaches, one of which involves hand-crafted features. We demonstrate through our experiments that hand-crafted features carry complementary discriminative information in addition to that present in the learned neural network models.