Convolutional Neural Networks for Sentiment Analysis on Amazon Product Reviews
In this paper, we compare several neural network architectures for analyzing the sentiments of Amazon product reviews (a.k.a. Product Reviews). Representative subset of data are selected from Baby product category collected in 2015. Two types of embeddings are experimented, namely, pre-trained GloVe and self-trained Word2Vec. We perform both binary and multiple sentiment labels classifications using a bag-of-word model (BoW) and proposed multi-resolution convolutional neural networks (MrCNNs). At the end we use commercial-off-the-shelf product (Amazon’s Comprehend service) to benchmark models’ performance in terms of accuracy.
Our unique contributions include the following. First, we propose a novel Mr-CNNs model. Second, we perform experiments on a few deep neural networks to demonstrate their performance in product review sentiment analysis. Finally our work focuses on the dataset from single natural language (i.e., English) and from the same business category (e.g. Baby product) to reduce complexity caused by languages and product terms.