Computational Social Science Approach to Social Media Analysis

Sentiment Analysis of Twitter Posts on Prime Minister Abe Shinzo from January 2020 to His Resignation

Published: March 1, 2023

Social media has a significant impact on society today, and it is inevitable to take into consideration the influence of social media in studying trends of public opinion. On the other hand, we need to examine what characteristics and challenges social media analysis has in comparison with the analysis of public opinion through traditional public opinion surveys. Nevertheless, analyzing discourse on social media platforms requires a different approach from conventional public opinion surveys. With this in mind, this paper discusses the characteristics of “public opinion” in the social media sphere by applying an emerging method in social science—a computational social science approach—to “public opinion” in social media to compare the results with those of a public opinion survey. As an example, Twitter posts on Prime Minister Abe Shinzo were analyzed.

To be specific, the method of sentiment analysis with supervised learning (SL) was used for the estimation of attitudes towards Mr. Abe, whether supporting him or not, from massive tweet data. As a machine learning model, we used RoBERTa, an improved model of BERT pre-trained language model based on deep learning. The model proved to function satisfactory with its accuracy rate of 85.79%.

The analysis finds that nearly 80% of tweets are classified as negative attitudes towards Mr. Abe, with support considerably exceeding disapprove throughout the observation period. Furthermore, we analyzed the words distinctive to the sentiments classified by the model, which confirmed that they were understandable to human eyes and the classification was adequate to some extent.

In this way, the chronological changes in support for and disapproval of Mr. Abe were compared with those of Cabinet approval ratings, which revealed that their results did not match and a huge gap was found. These findings will serve as good material for examining the relationship between opinions expressed on Twitter and general public opinion. In the next issue the paper will analyze in detail regarding what topics Mr. Abe was supported or disapproved of, using a method called topic model analysis, and discuss the usability and challenges of Twitter analysis.

The NHK Monthly Report on Broadcast Research

TAKIKAWA Hiroki / NAGAYOSHI Kikuko / LYU Zeyu / SHIMOKUBO Takuya / WATANABE Seiji / NAKAMURA Yoshiko

in Japanese