Characteristic Analysis of News Articles Based on Embedding Representations of Text and Graphs

Abstract

Embedding can be understood as a technique to map higher-order information into a lower-dimensional vector space. Currently, its applications span various domains such as images, natural language, and graphs. When obtaining embeddings for certain data, if they encompass various modalities, it can be expected to obtain embedding spaces with different nuances depending on each aspect. Therefore, in this study, we aim to generate embedding representations based on text obtained from news article titles and graphs obtained from reader networks, and attempt to understand article classification and characteristics using both representations. As a result, it was confirmed that there are articles with different browsing tendencies even if they are linguistically homogeneous, and conversely, articles that are linguistically heterogeneous but have the same browsing tendencies. Furthermore, in cases where articles are linguistically homogeneous but have different browsing tendencies, there are differences in sentiment polarity and readability compared to cases where browsing tendencies are the same, reflecting the possibility that the expression and description methods of media may differ even for articles with similar content.