Detecting Media Bias in News Articles using Gaussian Bias Distributions

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4290–4300 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 4290 Detecting Media Bias in News Articles using Gaussian Bias Distributions Wei-Fan Chen Paderborn University Department of Computer Science cwf@mail.upb.de Khalid Al-Khatib Bauhaus-Universität Weimar Faculty of Media, Webis Group khalid.alkhatib@uni-weimar.de Benno Stein Bauhaus-Universität Weimar Faculty of Media, Webis Group benno.stein@uni-weimar.de Henning Wachsmuth Paderborn University Department of Computer Science henningw@upb.de Abstract Media plays an important role in shaping pub- lic opinion. Biased media can influence peo- ple in undesirable directions and hence should be unmasked as such. We observe that feature- based and neural text classification approaches which rely only on the distribution of low-level lexical information fail to detect media bias. This weakness becomes most noticeable for ar- ticles on new events, where words appear in new contexts and hence their “bias predictive- ness” is unclear. In this paper, we therefore study how second-order information about bi- ased statements in an article helps to improve detection effectiveness. In particular, we uti- lize the probability distributions of the fre- quency, positions, and sequential order of lex- ical and informational sentence-level bias in a Gaussian Mixture Model. On an existing me- dia bias dataset, we find that the frequency and positions of biased statements strongly impact article-level bias, whereas their exact sequen- tial order is secondary. Using a standard model for sentence-level bias detection, we provide empirical evidence that article-level bias detec- tors that use second-order information clearly outperform those without. 1 Introduction Media bias is discussed and analyzed in journalism research (Groseclose and Milyo, 2005; DellaVi- gna and Kaplan, 2007; Iyengar and Hahn, 2009) and NLP research (Gerrish and Blei, 2011; Iyyer et al., 2014; Chen et al., 2018). According to the study of Groseclose and Milyo (2005), bias “has nothing to do with the honesty or accuracy”, but it means “taste or preference”. In fact, journalists may (1) report facts only in favor of one particular political side and thus (2) conclude with their own opinion. As an example, the following sentences from allsides.com reporting on the event “Trump asks if disinfectant, sunlight can treat coronavirus” demonstrate media bias on the sentence level: The activists falsely claimed that Trump “urged Americans to inject themselves with disinfectant” and “told people to drink bleach.” — The Daily Wire, right-oriented Lysol maker issues warning against injections of disinfectant after Trump comments — The Hill, center-oriented “This notion of injecting or ingesting any type of cleansing product into the body is irresponsible and it’s dangerous,” said Gupta. — NBC News, left-oriented From an NLP perspective, bias in the example sentences could be detected by capturing sentiment words, such as “falsely” or “irresponsible”. With- out the background knowledge of the political side of Trump or the event itself, however, predicting which side these sentences are slanted to is difficult. Bias detection even becomes harder at the article level. For illustration, Figure 1 shows two articles and their sentence-level bias from the used dataset. It …

Content truncated to 500 words