Content Analysis Based Review Organization Methods

The rapid growth in E-Commerce, as well as its support on information interaction, brings huge user-generated reviews. A product review consists of a numeric rating and a piece of unstructured textual description. The numeric rating is an overall evaluation on the comment product, while the review text provides detailed information on product at-tributes. Review content reflects users’opinions on products they bought or services they experience, so it is one of the keys to user consumption. Merchants try to build reputations and discover problems through reviews; consumers read reviews to get a clear picture of products. Therefore, it becomes an important part of E-Commerce sites to collect, pre-process and analyze review data. However, with the popularity of mobile applications and wireless networks, review data proliferate in recent years, breaking through the limitation-s of time and space. Moreover, due to the emergence of mobile devices, users are in need of a concise but comprehensive overview of review content, which makes it more difficult to process review data. As there exists a large volume of reviews in E-Commerce sites, it is critical to analyze and organize review content, thus extracting useful information from review data.Effectively organizing review content not only solves the problem of information overload, but also improves user experience, thus enhancing the usability of E-Commerce sites. Current review organization methods rank and return top-K reviews according to review quality. Review quality can be assessed from the coverage of product attributes and the consistency in opinion distributions. However, those methods neither measure the importance of different attributes, nor consider the diversity of opinions in the top-K reviews. Therefore, an attribute importance based approach is proposed to select a rep-resentative set of reviews for each product, which is of high quality and diversified in opinions. In addition, it is common in existing work that each review is associated with one product, i.e., each review has one comment target. When applied to review data such as restaurant reviews, the above representative approach has a drawback:coarse grain. A single restaurant review may comment on several dishes, and users prefer dish-oriented information via review organization. To generate summarization for each product men-tioned in a review set, an approach for reviews containing multiple comment targets is proposed. Product-oriented summarization, including an overall rating and K represen-tative review snippets, is generated for each product in the review set.In conclusion, the main contributions of this paper are as follows:· Review Quality Assessment and Top-K Review Selection:1. First, a method is proposed to measure the importance of each product at-tribute and assign a weight to it. Considering the importance of attributes can improve the accuracy of review quality assessment.2. Second, a cluster-based method is introduced to measure the diversity of re-views, in which reviews are clustered according to their distributions on at-tributes and opinions.3. Then, a diversified algorithm is designed to select reviews proportionally from each cluster. The top-K reviews can cover more product attributes and reflect the opinion distributions in the original review set.4. Finally, experiments on product review data crawled from online shopping sites show the effectiveness of the proposed top-K review assessment and selection approach for reviews commenting on one target.· Product-Oriented Review Summarization:1. First, a method is proposed to find whether a review snippet contains user opinion. The method is unsupervised and classifies review snippets according to the information entropy of each N-gram within.2. Second, three methods are introduced to extract the opinion of a review snip-pet, in which the overall ratings of reviews are used to predict the rating of the review snippet as its opinion. A product rating can be further calculated based on the ratings of snippets for the product.3. Then, two diversified algorithms are designed to select review snippets propor-tionally from each cluster that is either opinion-based or attribute and opinion based. The K review snippets can cover a diversified set of product attributes and user opinions from the original review set.4. Finally, experiments on restaurant review data crawled from online review sites show the effectiveness of the proposed product-oriented review summa-rization approach for reviews commenting on several targets.