Classification Model Learning for Bulletin Board Site Analysis Based on Unbalanced Textual Examples

Abstract

This paper proposes a method that acquires a more appropriate classification model for label extraction. The model can extract specific labels from articles included in bulletin board sites. The labels represent the contents of the articles and are used to characterize the articles. The method selects two kinds of important examples not including a specific label by using expressions related to the label. The method inductively acquires the classification model from the selected examples and examples including the label. The paper applies the method to articles collected from three bulletin board sites and verifies its effect through comparative experiments.