Learning Extraction of Chinese Comparative Sentences for Evaluative Text

With the prevalence of Web 2.0, people increasingly prefer to express opinions and exchange information through CGM (consumer-generated media), such as blog, Internet forum and etc. Many studies pay attention to extract and analysis user opinions in consumer reviews. This paper studies how to automatically extract Chinese comparative sentences from consumer reviews. At first, the paper describes a method for solving the class imbalance problem of comparatives and non-comparatives in review data. Then we built a support vector machine learning model to classify comparatives and noncomparatives into different group on a balanced dataset. Experiments were conducted on consumer-generated product reviews, including 9600 sentences, of which 1,624 (16.92% of the total) were comparisons. Experiments show an overall F-score of 87.26%, which presents the effectiveness of the proposed approach.