Investigation of the Diversity of News Reading Article and Browsing Trends Using Sentence-BERT
Akihisa Takiguchi, Tsunenori Mine, Yutaka Arakawa · IIAI Letters on Informatics and Interdisciplinary Research · 2024
In modern life, platforms like a social networking service (SNS) play a crucial role by utilizing recommender systems to present useful information from vast datasets. However, the advancement of these systems has led to biases in user-exposed information, causing societal issues like public opinion conflicts and defamation. Furthermore, sentiment biases in the information viewed contribute to this problem, described as “informational malnutrition”. This highlights the need for “informational health”, where user access to various information maintains the balance of information intake they seek. In this work, we explored differences in user viewing tendencies based on the diversity of viewed articles, employing a dataset of news articles and user logs. We utilized Sentence-BERT, a natural language processing model known for its effective sentence similarity analysis, to vectorize articles and score their similarity, measuring users’ article diversity. Our analysis, considering sentiment content biases, used multiple regression. The results suggest that users with diverse viewing habits tend to prefer articles with a negative bias and that news in categories such as music, current affairs, and politics have a low contribution to the diversity of information viewed, and conversely, categories like entertainment and lifestyle content tend to have a high contribution to the diversity of information viewed.