Discovering Opinion Changes in Online Reviews via Learning Fine-Grained Sentiments

Abstract

Recent years have shown rapid advancement in understanding consumers' behaviors and opinions through collecting and analyzing data from online social media platforms. While abundant research has been undertaken to detect users' opinions, few tools are available for understanding events where user opinions change drastically. In this paper, we propose a novel framework for discovering consumer opinion changing events. To detect subtle opinion changes over time, we first develop a novel fine-grained sentiment classification method by leveraging word embedding and convolutional neural networks. The method learns sentiment-enhanced word embedding, both for words and phrases, to capture their corresponding syntactic, semantic, and sentimental characteristics. We then propose an opinion shift detection algorithm that is based on the Kullback-Leibler divergence of temporal opinion category distributions, and conducted experiments on online reviews from Yelp. The results show that the proposed approach can effectively classify fine-grained sentiments of reviews and can discover key moments that correspond to consumer opinion shifts in response to events that relate to a product or service.