Public opinion prediction with language models.
Media plays an important role in shaping people's beliefs and behaviors.
Mass media and social media reflect and form public opinion, which can ultimately lead to positive outcomes such as civil engagement with opposing tribes, but also negative outcomes such as non-adherence to health-beneficial social guidelines. Understanding viewpoints expressed online and the relationship between media and beliefs is increasingly pertinent today, in a world of constant connectivity and "alternate realities". Motivated by the effects of misinformation in the COVID-19 pandemic, we introduce a new approach around "media diet models", which can act as proxies for human media consumption.
By probing these models, we can predict public opinion as measured by nationally representative surveys with media diet information for each respondent. We examine this approach on surveys conducted during the COVID-19 pandemic, as well as monthly consumer confidence surveys.
In both settings, we find there to be predictive power in the media diet models.
Particularly in the COVID-19 setting, there is strong correlation between the probed responses and survey responses, and evidence such a model can act as a leading indicator of public opinion. Finally, we show several ways to scale this approach to automatically generate insights around the heterogeneity of messaging and possible induced human beliefs.