Role of Data Properties on Sentiment Analysis of Texts via Convolutions

Dense and low dimensional word embeddings opened up the possibility to analyze text polarity with highly successful deep learning techniques like Convolution Neural Networks. In this paper we utilize pretrained word vectors in combination with simple neural networks of stacked convolution and max-pooling layers, to explore the role of dataset size and document length in sentiment polarity prediction.We experiment with song lyrics and reviews of products or movies and see that convolution-pooling combination is very fast and yet quiet effective. We also find interesting relations between dataset size, text length and length of feature maps with classification accuracy. Our next goal is the design of a generic neural architecture for analyzing polarity of various text types, with high accuracy and few hyper-parameter changes.