Aspect-based Review Extraction for E-Commerce Products

Nowadays, more and more products are sold online. Under popular products, there are normally hundreds or even thousands of reviews left by the previous customers. These reviews help potential buyers understand the products better and make the purchase decision. However, most shopping websites only give an overview score (e.g. 3 star out of 5) of a product besides the reviews, which does not provide enough information for people to understand different aspects of the product. For example, if the users want to know more about the ``picture quality'' aspect of a camera product, they need to read the reviews. However, by reading the top ranked reviews or some random reviews, they may get biased information. Professional people may have high demands in ``picture quality'' while the top ranked reviews may be amateurs who are easy to get satisfied. However, it is tedious for the users to read all of the reviews to get the right information. Moreover, customers tend to have different opinions on one product, but the current shopping websites do not cluster the reviews which share similar opinions together and present it to the users. Instead, they ignore the conflicted opinions in the reviews by simply averaging the scores given by all the reviewers. Under this circumstance, this thesis proposes to research questions: is it possible to automatically extract aspects of online products which are also consistent with the manual modeling of products in this domain? Is it possible to build a system to aggregate opinions on online products from reviews, which actually can help users understand the products faster? To answer the two research questions, it is necessary to build a prototype system which can extract information and describe different aspects of a product to help users understand the product. This thesis first develops such a system. The system is divided into three steps, aspect detection, sentiment analysis and reviews clustering. In brief, the system first determines which aspects should be used to describe the products, and then calculates people's sentiment towards these aspects, in the end the system clusters reviews which share similar opinions together. To answer the research questions, two evaluations of the system are conducted in the end, which shows the system has big potential in helping users understanding online shopping products