Shedding (a Thousand Points of) Light on Biased Language Tae Yano Philip Resnik School of Computer Science Department of Linguistics and UMIACS Carnegie Mellon University University of Maryland Pittsburgh, PA 15213, USA College Park, MD 20742, USA
[email protected] [email protected] Noah A. Smith School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA
[email protected] Abstract To what extent a sentence or clause is biased (none, • somewhat, very); This paper considers the linguistic indicators of bias in political text. We used Amazon Mechanical Turk The nature of the bias (very liberal, moderately lib- • judgments about sentences from American political eral, moderately conservative, very conservative, bi- blogs, asking annotators to indicate whether a sen- ased but not sure which direction); and tence showed bias, and if so, in which political di- rection and through which word tokens. We also Which words in the sentence give away the author’s asked annotators questions about their own political • views. We conducted a preliminary analysis of the bias, similar to “rationale” annotations in Zaidan et data, exploring how different groups perceive bias in al. (2007). different blogs, and showing some lexical indicators strongly associated with perceived bias. For example, a participant might identify a moderate liberal bias in this sentence, 1 Introduction Without Sestak’s challenge, we would have Bias and framing are central topics in the study of com- Specter, comfortably ensconced as a Democrat munications, media, and political discourse (Scheufele, in name only. 1999; Entman, 2007), but they have received relatively adding checkmarks on the underlined words.