Calibrated and Diverse News Coverage

Abstract

In recent years, there has been a debate about whether automated news aggregators, like Google News, lead readers to content that reinforces their existing beliefs and restricts their exposure to a biased subset of perspectives. To avoid bias, it has become common practice that news aggregators provide articles based on source diversity: for each story, they pick articles from news sources with different political leanings. In this paper, we ask whether this practice is sufficient. In particular, we study how well the diversity of viewpoints, in particular with respect to entities, is covered by articles picked using plain source diversity. We analyze a dataset fetched from Google News and find that, even though the top articles exhibit some diversity with respect to the leanings of the news outlets, many possible viewpoints towards the entities are missing. Based on this observation we design novel methods for selecting a small set of articles that cover all possible viewpoints; to ensure that our selections are useful we show how to incorporate the user preferences into our model. Our experiments on four real-world datasets show that our algorithms cover significantly more different viewpoints than previous baselines.