Present-day media companies produce an enormous amount of news content. The New York Times publishes roughly 1.100 articles each day, and not even the most dedicated reader can keep up with all this. Users rely on news recommenders to pick the articles that are most relevant and interesting to them. Because of this, news recommenders influence the information we receive about the world and therefore indirectly shape our view of it. But how can we know if this does not place us in a so-called filter bubble, confirming our existing views of the world? And how can media companies that build recommendation systems that avoid the risks that come with collaborative filtering and optimizing on user clicks?

In a research project in collaboration with RTL Nieuws, the University of Amsterdam is working on a tool that measures generated recommendations according to different interpretations of diversity, ranging from the field of computer science to democratic theory. This tool is built using only open source software including Elasticsearch, SpaCy and Wikidata, and is publicly accessible. In this talk, we will explain the necessity of building a tool like this, and the steps undertaken to develop a tool that implements the different metrics of diversity. We discuss the difficulties and advantages of working in a multi-disciplinary setting, and aim to stimulate discussion and thought about the way we construct recommendation systems.

Project repository: https://github.com/svrijenhoek/dart (work still ongoing, to be completed September 2019).