Linguistic Indicators of Trust in Written News

Americans’ trust and confidence in the mass media is at an all-time low. Compounding this effect is the increase in “fake news.” It is becoming increasingly difficult for media consumers to distinguish between truthful and fictional news stories; readers of fake news reports often believe them, while readers of accurate reports question and mistrust them. A free press is a critical component of democracy, yet it seems to be in danger in the current age of media mistrust. To combat this trend, there have been recent efforts in the Natural Language Processing (NLP) community to use machine learning to automatically distinguish between “real” and “fake” news. This work is very important and will hopefully equip media consumers with the necessary tools to navigate the murky world of truth in media.

This project aims to study a complementary problem to fake news detection: trusted news detection. Instead of focusing on determining what is true or fictional, this study aims to discover the characteristics of trusted or believed text, regardless of the veracity of the text in question. Trust in media has been previously studied qualitatively: this proposed research is to our knowledge the first effort to quantitatively study trust in media on a large scale, using automated crowdsourcing, machine learning, and natural language processing methods. Further, this project proposes to analyze group-specific indicators of trust, to discover whether perception of trustworthiness varies across different categories of media consumers.

Project leads: Julia Hirschberg, Sarah Ita