VizByWiki: Mining Data Visualizations from the Web to Enrich News Articles
Allen Yilun Lin, Joshua Ford, Eytan Adar, and Brent Hecht
Data visualizations in news articles (e.g., maps, line graphs, bar charts) greatly enrich the content of news articles and result in well-established improvements to reader comprehension. However, existing systems that generate news data visualizations either require substantial manual effort or are limited to very specific types of data visualizations, thereby greatly restricting the number of news articles that can be enhanced. To address this issue, we define a new problem: given a news article, retrieve relevant visualizations that already exist on the web. We show that this problem is tractable through a new system, VizByWiki, that mines contextually relevant data visualizations from Wikimedia Commons, the central file repository for Wikipedia. Using a novel ground truth dataset, we show that VizByWiki can successfully augment as many as 48% of popular online news articles with news visualizations. We also demonstrate that VizByWiki can automatically rank visualizations according to their usefulness with reasonable accuracy (nDCG@5 of 0.82). To facilitate further advances on our "news visualization retrieval problem", we release our ground truth dataset and make our system and its source code publicly available.
Preprint: PDF (1.2Mb), To appear at WWW'18