Visual Insights: Enhancing Online News Maps with Relevant Illustrations

In recent years, documents such as news articles, blog articles, scientific papers are posted on the Internet every moment, and techniques for processing a huge amount of text data have been proposed. The development of natural language processing technology has made it easier to acquire embedding vectors that express the meaning of words and sentences, and they are used as inputs for various machine learning methods. In addition, a technique has been established to compress high-dimensional embedding vectors into low-dimensional ones, which makes it possible to represent a group of documents as a two-dimensional map. However, since it is not obvious where on the map what kind of documents are gathered, it is necessary to add information in some way. In this study, we propose a method to enhance the understanding of the overall picture of current events and enable cross-category exploration of related articles by visualizing online news articles on a two-dimensional map. Online news articles are a valuable source of information and are accessed by many users. To achieve our goal, we annotate the news map with relevant illustrations, making it a useful tool for users of various age groups and various nationality groups to grasp the broader context of events. Our approach involves embedding the distributed representations of news articles into a two-dimensional space, dividing the space into regions where related articles cluster, and extracting significant and frequently occurring words within each region. We then measure the similarity between the distributed representations of these words and illustrations to annotate relevant illustrations to the corresponding regions. Through evaluation experiments using real data, both quantitatively and qualitatively, we demonstrate the effectiveness of our proposed method, which utilizes Doc2Vec, a technique for embedding distributed representations.