Discovering Attribute-Specific Features From Online Reviews

Abstract

This article describes how online reviews play an important role in data driven decision making. Many efforts have been invested in determining the overall sentiment carried by the reviews. However, oftentimes, the overall ratings of the reviews do not represent opinions toward specific attributes of a product. An ideal opinion mining tool should aim at finding both the product attributes and their corresponding opinions. The authors propose an approach for extracting the attribute specific features from online reviews using a Word2Vec model combined with clustering. Two experiments are described in this paper: the first focuses on testing the performance of the Word2Vec model on extracting product aspect words, the second addresses how well the extracted features obtained are recognizable by human cognition. A new metric called the “split value” that is based on cluster similarity and diversity is introduced to examine the consistency of clustering algorithm. The authors' experiments suggest that meaningful clusters, which provide insights to the product attributes and sentiments, could be extracted from the reviews.