Improving a method for quantifying readers' impressions of news articles with a regression equation

  • Authors:

  • Tadahiko Kumamoto;Yukiko Kawai;Katsumi Tanaka

  • Affiliations:

  • Chiba Institute of Technology, Tsudanuma, Narashino, Chiba, Japan;Kyoto Sangyo University, Motoyama, Kamigamo, Kita-Ku, Kyoto, Japan;Kyoto University, Yoshida-Honmachi, Sakyo-Ku, Kyoto, Japan

  • Venue:
  • WASSA '11 Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sentiment Analysis
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

In this paper, we focus on the impressions that people gain from reading articles in Japanese newspapers, and we propose a method for extracting and quantifying these impressions in real numbers. The target impressions are limited to those represented by three bipolar scales, "Happy -- Sad," "Glad -- Angry," and "Peaceful -- Strained," and the strength of each impression is computed as a real number between 1 and 7. First, we implement a method for computing impression values of articles using an impression lexicon. This lexicon represents a correlation between the words appearing in articles and the influence of these words on the readers' impressions, and is created from a newspaper database using a word co-occurrence based method. We considered that some gaps would occur between values computed by such an unsupervised method and those judged by the readers, and we conducted experiments with 900 subjects to identify what gaps actually occurred. Consequently, we propose a new approach that uses regression equations to correct impression values computed by the method. Our investigation shows that accuracy is improved by a range of 23.2% to 42.7% by using regression equations.