Analysis of Sentiment and Stance Bias of News Recommender Systems

What is it about?

News recommender systems are increasingly used by online news providers to alleviate the users' information overload by providing them with customized suggestions. Increased personalization, however, is hypothesized to reduce the users' exposure to diverse content and to result in filter bubbles. This paper uses sentiment and stance information extracted from German news articles to examine and quantify the extent to which recommender systems are prone towards particular sentiments or stances expressed in the news. More specifically, we evaluate the behavior of four different recommendation models and report on their (a) tendency towards different sentiments and stances (b) predictive performance (c) interplay between accuracy and diversity of recommendations.

Featured Image

Photo by Rudolf-Peter Bakker on Unsplash

Photo by Rudolf-Peter Bakker on Unsplash

Why is it important?

Online news consumption is largely influenced by recommender systems which curate a news selection tailored to the users' preferences. Consequently, algorithmic news curation shapes the users' perception of the world by influencing their exposure to diverse content -- articles deemed irrelevant or inconsistent with the users' opinions are often filtered out to increase user engagement. In the context of news, an over-exposure to similar ideological viewpoints could, in the long run, lead to opinion polarization. Therefore, it is critical to study the effects of recommender systems in the creation of filter bubbles, and to understand how these could be mitigated to achieve a more balanced and diverse news consumption. While the theory of filter bubbles and polarization is rather elaborate, such quantitative studies are rather rare, and this work constitutes a step further at closing the gap.

Perspectives