Sentiment analysis via Doc2Vec and Convolutional Neural Network hybrids

In this study, we proposed two ensembled Convolutional Neural Network architectures viz. (CNNcuPSONN) and CNN-PNN, where cuPSONN is a CUDA enabled particle swarm optimization optimized neural network and PNN is the probabilistic neural network. We compared their performance with that of the standalone PNN. The techniques are invoked after distributed Doc2Vec was employed on movie review dataset of the size of 992 MB, to get the paragraph embeddings of the dataset. Apache Spark framework was used to get the embeddings. Here, we considered CNN as a feature extractor from the embeddings, while cuPSONN and PNN perform classification. The third proposed model is PNN preceded by t-statistic based feature selection (t-statistic-PNN). Among the three classifiers employed, CNN-PNN yielded an area under ROC curve of 96.63% under the 10fold cross validation framework. The proposed CNN-PNN turned out to be statistically significant with respect to the CNN-cuPSONN and t-statistic-PNN, while it is statistically the same as the distributed multi-layer perceptron (DMLP), which is considered as baseline classifier to compare the performance of our proposed algorithms.