A METHOD OF EXTRACTING AND EVALUATING GOOD AND BAD REPUTATIONS FOR NATURAL LANGUAGE EXPRESSIONS

Abstract

Although a users' opinion or a live voice is a very useful information for text mining of the business, it is difficult to extract good and bad reputations of users from texts written in natural language. The good and bad reputations discussed here depend on users' claims, interests and demands. This paper presents a method of determining these reputations in commodity review sentences. Multi-attribute rule is introduced to extract the reputations from sentences, and four-stage-rules are defined in order to evaluate good and bad reputations step by step. A deterministic multi-attribute pattern matching algorithm is utilized to determine the reputations efficiently. From simulation results for 2,240 review comments, it is verified that the multi-attribute pattern matching algorithm is 63.1 times faster than the Aho and Corasick method. The precision and recall of extracted reputations for each commodity are 94% and 93% respectively. Moreover, the precision and recall of the resulting reputations for each rule are 95% and 95% respectively.