Reputation Analysis Using Key Phrases and Sentiment Scores Extracted from Reviews

Reputation Analysis Using Key Phrases and Sentiment Scores Extracted from Reviews Huang Yipu Graduate School of Science and Engineering, Ibaraki University Minoru Sasaki Computer and Information Sciences, Faculty of Engineer- ing, Ibaraki University Kanako Komiya Graduate School of Engineer- ing, Tokyo University of Ag- riculture and Technology {21nm723s, minoru.sasaki.01}@vc.ibaraki.ac.jp, kkomiya@go.tuat.ac.jp Abstract In recent years, reviews of various products by purchasers have been posted on websites. By browsing these reviews, company representa- tives intend to collect and analyze opinions from the users' side of the products and use them to improve the products. However, exist- ing analysis methods analyze reviews by divid- ing them into words, and do not analyze the function names that are phrases. Therefore, we propose a reputation analysis method using a decision tree model based on the frequency of key phrases extracted from reviews and senti- ment scores. Experimental results showed that the proposed method using key phrases im- proved the accuracy by 3 points compared to the existing method that analyzes by words. 1 Introduction When consumers purchase a product or service, they invariably consult reviews. In fact, review analysis has many benefits not only for individuals but also for companies． For example, the use of the analysis for product planning, analysis of user needs and dissatisfaction, hints for function and service improvement, and verification of the effectiveness of promotions and marketing measures are the effects of corporate re- view analysis. In order to improve a product, it is necessary to know what reviewers like about the product and how they evaluate it. Therefore, we use two criteria, the frequency of occurrence of key phrases and the sentiment score, to determine the reviewer's opinion of the product. For example, if the frequency of oc- currence of a key phrase is high and the sentiment score is also high, we judge that "everyone is inter- ested in the product, and it has a good reputation" and that the product does not need to be improved immediately. This time, key phrases (important words) are ex- tracted from a certain review, and a dictionary is created with key phrases as keys and the frequency of occurrence of key phrases as values. Then, we calculate the sentiment score from the review sen- tences, treat the sentiment score and frequency of key phrases as explanatory variables, and treat whether improvement is necessary or not or priority as objective variables, and create a decision tree model. The following sections describe key phrase ex- traction, sentiment score calculation, decision tree model creation, and quality evaluation. 2 Related Research There have been several studies on review analysis methods. Kobayashi collects and analyzes reviews from web pages, they use TF-IDF to extract key- words, and then extract emotion words, and com- bine keywords and emotion words (Kobayashi, 2008). However, it is difficult to grasp the whole picture because of the variety of emotion words ex- tracted from various reviews. Abe analyzes product reviews, assigns a score to each evaluation item, and proposes to create an …

Content truncated to 500 words