Sentiment Analysis Based on a Deep Stochastic Network and Active Learning

Abstract

this paper proposes a novel approach for sentiment analysis. The growing importance of sentiment analysis commensurate with the use of social media such as reviews, forum discussions, blogs, micro blogs like Twitter, and other social networks. We require efficient and higher accuracy algorithms in sentiment polarity classification as well as sentiment strength detection. In comparison to pure vocabulary based system, deep learning algorithms show significantly higher performance. The goal of this research is to modify a Recurrent Neural Network (RNN) with Gated Recurrent Unit (GRU) by introducing stochastic depth in a hidden layer and comparing it with baseline Naïve Bayes, vanilla RNN and GRU-RNN models. To improve our results, we also incorporated Active Learning with Uncertainty Sampling approach. Movie review dataset from Rotten Tomatoes was used, the dataset includes 215,154 fine grained labelled phrases in addition to 11,855 full sentences. We performed pre-processing on the data and used an embedding matrix with pre-trained word vectors as features for training our model. These word vectors were generated using character level n-grams with fasttext on Wikipedia data.