Table of Links

Abstract and Introduction

Related Work

Study 1: Evolution of Semantic Polarity in Broadcast Media Language (2010-2020)

Study 2: Words that Characterize Semantic Polarity between Fox News & CNN in 2020

Study 3: How Semantic Polarization in Broadcast Media Language Forecasts Semantic Polarity in Social Media Discourse

Discussion and Ethics Statement

Appendix and References

Discussion

The rising levels of semantic polarity between the two major broadcast news organizations, as demonstrated in our work may render people’s ability to reach across partisan divides and to perceive and solve issues democratically more difficult. The way CNN and Fox News discuss identical keywords on their programs is remarkably distinct in 2020 where semantic polarity between the two stations reaches its peak over an 11-year period, corroborating the widening partisan media gap highlighted by recent scholarship. Framing Theory argues that even subtle changes in the wording around how an issue is described by the media can significantly influence how audiences interpret the issue (Scheufele 2000; Entman 2003). Yet, our findings show that the contextual language driving semantic polarization in broadcast media is not nuanced at all. The surrounding words that characterize how each station discusses topically important keywords are drastically different, to the extent that identical words seem to reflect different meanings altogether. This is a crucial point of consideration, as linguistic cues in the media can play a powerful role in selectively reifying certain aspects of the perceived reality of an issue while downplaying others. Our findings suggest that TV news language does not only shape how online audiences perceive issues, but also how they talk about them.

In linguistics, collocation is the property of language where two or more words appear in each other’s company with greater than random chance (Hoey 2005), such as “illegal” and “immigrants” in Fox News and “climate change” and “science” from CNN, as shown in our findings. Repeated encounters with collocated words drive what linguists call lexical priming (Hoey 2005), in which a priming word (e.g., ”blacklivesmatter”) provokes a particular target word (“protests” for CNN and “violence” for Fox) more strongly and quickly than disparate words encountered more rarely. In this vein, online audiences who consume very different perceived realities from the media may be lexically primed through repeated exposure to collocated words that frame identical issues in acutely contrasting contexts. This may help theorize why TV news language can shape how online audiences engage in public discourse as demonstrated in this work. Semantic polarization in televised media not only forecasts semantic polarity trends on Twitter, but the words that characterize broadcast media polarization re-appear across Twitter discussions, separated by significant lag months. Our results demonstrate that the language characterizing opposing media narratives around topically similar news events on TV can be linguistically mimicked in how social media users are polarized in their discourse around important social issues.

Limitations. Differences in cosine values between contextual word representations stem not only from semantic, but also from positional differences of words in sentences. Although word embeddings generally embody contextual nuances, we acknowledge that our work predominantly considers semantic rather than syntactic aspects, and that to some extent our calculation of semantic polarity could embody syntactic differences in how keywords are discussed. Furthermore, our findings pertain to American political contexts, which might not be generalizable to foreign public discourse and media language. For future work, we aim to apply our model to relevant textual corpora in other languages by incorporating cross-lingual transfer learning methods to better understand the generalizability of our model by comparing distributional differences with cross-lingual data.

Ethics Statement

We aim to strictly adhere to the AAAI Code of Ethics and Conduct by ensuring that no personal information of Twitter users was collected nor compromised throughout our project. All data in this work are securely stored on servers only accessible by the authors. Our semantic polarization framework is made publicly available on the authors’ GitHub page and can be applied on a variety of textual corpora with diachronic and contrastive properties.