SentiLDA — An Effective and Scalable Approach to Mine Opinions of Consumer Reviews by Utilizing Both Structured and Unstructured Data

With the help of Internet and Web technologies, more and more consumers tend to seek opinions online before making purchase decisions. However, with the ever-increasing volume of user generated reviews, people are overwhelmed with the amount of data they have. Thus there is a great need for a system that can summarize the reviews and produce a set of aspects being mentioned in the reviews together with the pros/cons being expressed to them. To address the need, this paper proposes a new probabilistic topic model, SentiLDA, for mining reviews (unstructured data) and their ratings (structured data) jointly to detect the product/service aspects and their corresponding positive and negative opinions simultaneously. A key feature of SentiLDA is that it is capable of mining positive and negative sub-topics under the same aspect without the need of sentiment seed words. Experiment results show that the performance of SentiLDA outperforms the other related state-of-the-art models in detecting product/service aspects and their corresponding sentiments in reviews.