Understanding Characteristics of Biased Sentences in News Articles

Providing balanced and good quality news articles to readers is an important challenge in news recommendation. Often, readers tend to select and read articles which confirm their social environment and their political beliefs. This issue is also known as filter bubble. As a remedy, initial approaches towards automatically detecting bias in news articles have been developed. Obtaining a suitable ground truth for such a task is however difficult. In this paper, we describe ground truth dataset created with the help of crowd-sourcing for fostering research on bias detection and removal from news content. We then analyze the characteristics of the user annotations, in particular concerning bias-inducing words. Our results indicate that determining bias-induced words is subjective to certain degree and that a high agreement on all bias-inducing words of all readers is hard to obtain. We also study the discriminative characteristics of biased content and find that linguistic features, such as negative words, tend to be indicative for bias.