MRR

Abstract

The automatic detection of relevant reviews plays a major role in tasks such as opinion summarization, opinion-based recommendation, and opinion retrieval. Supervised approaches for ranking reviews by relevance rely on the existence of a significant, domain-dependent training data set. In this work, we propose MRR (Most Relevant Reviews), a new unsupervised algorithm that identifies relevant revisions based on the concept of graph centrality. The intuition behind MRR is that central reviews highlight aspects of a product that many other reviews frequently mention, with similar opinions, as expressed in terms of ratings. MRR constructs a graph where nodes represent reviews, which are connected by edges when a minimum similarity between a pair of reviews is observed, and then employs PageRank to compute the centrality. The minimum similarity is graph-specific, and takes into account how reviews are written in specific domains. The similarity function does not require extensive pre-processing, thus reducing the computational cost. Using reviews from books and electronics products, our approach has outperformed the two unsupervised baselines and shown a comparable performance with two supervised regression models in a specific setting. MRR has also achieved a significantly superior run-time performance in a comparison with the unsupervised baselines.