Towards opinion mining from reviews for the prediction of product rankings

Abstract

Opinion mining aims at summarizing the content of reviews for a specific brand, product, or manufacturer. However, the actual desire of a user is often one step further: Produce a ranking corresponding to specific needs such that a selection process is supported. In this work, we aim towards closing this gap. We present the task to rank products based on sentiment information and discuss necessary steps towards addressing this task. This includes, on the one hand, the identification of gold rankings as a fundament for an objective function and evaluation and, on the other hand, methods to rank products based on review information. To demonstrate early results on that task, we employ real world examples of rankings as gold standard that are of interest to potential customers as well as product managers, in our case the sales ranking provided by Amazon.com and the quality ranking by Snapsort.com. As baseline methods, we use the average star ratings and review frequencies. Our best text-based approximation of the sales ranking achieves a Spearman's correlation coefficient of ρ = 0.23. On the Snapsort data, a ranking based on extracting comparisons leads to ρ = 0.51. In addition, we show that aspect-specific rankings can be used to measure the impact of specific aspects on the ranking.