Seeing Stars from Reviews by a Semantic-based Approach with MapReduce Implementation

This study concerns the problem of aspect-level opinion (sentiment) mining from online reviews. The problem consists of two fundamental sub-tasks: aspect extraction (identify specific aspects of the product from reviews), and aspect rating estimation (offer a numerical rating for each aspect). Solving this problem is important and useful for many applications, e.g., providing aspect-level review summaries to consumers for better decision making, and for product manufacturers to collect summarized user feedback. Our objective is to propose a semantic-based approach for aspect level opinion mining from massive amounts of reviews in a scalable fashion. The MapReduce implementation for this approach obtains much runtime reduction compared with the single-process implementation. Experimental results show that the runtime reductions by the MapReduce implementation are almost linear to the number of mappers, e.g., around 7.4 times reduction with 10 mappers on the TripAdvisor dataset and 2.6 times reduction with 4 mappers on the Yelp dataset. The number of mappers and reducers can be configured on demand to handle very large datasets in a scalable fashion. Moreover, the semantic-based approach obtains good performance for aspect rating estimation on the TripAdvisor dataset, with the MAE score of around 1.0 on all aspects, which means that the average deviation between the human rating and the estimated rating is around 1 star. The source code of our implementation for the sentiment-based approach can be downloaded from https://github.com/ppfliu/aspect-opinion.