Going Big and Deep: Using Convolutional Neural Network to Leverage Training Data from Multiple Domains for Cross-Domain Sentiment Classification on Product Reviews

Abstract

Training a classifier for sentiment polarity detection in product reviews when labeled data is not available for a particular domain poses a challenge, which can be addressed through cross-domain sentiment analysis. We experimented with Convolutional Neural Network (CNN) to learn sentiment polarity (positive or negative) from labeled data available in many different source domains and test its performance on a target domain that it is not trained on. Extensive experiments were conducted on 14 different domains using Amazon product reviews. Our preliminary findings show that cross-domain CNN models trained with multiple source domains achieved accuracy of above 80% across all the domains and outperform the in-domain models trained using limited labeled data from the same domain. In fact, the cross-domain CNN models demonstrated better performance when a larger number of source domains are used for training. Therefore, going deep and big is a promising direction to explore for cross-domain sentiment classification.