Understanding media narratives with machine learning and NLP

Prof. Lu Wang aims to build a better understanding of how storytelling in mass media impacts the attitudes and beliefs of individuals and communities.
Person holding a phone with the news app open

Storytelling and narrative crafting are central to communication techniques — so much so that they drive the way news media, advertising, and public relations operate today. But the way narratives are used in these communications, as well as how they impact the opinions of individuals or an entire society, is extremely complex and difficult to express with any specificity. A new project at the University of Michigan supported by the Air Force Office of Scientific Research (AFOSR) aims to use computational tools to conceptualize these narratives and the impact they have on readers.

“It remains unclear how to effectively represent and extract narratives at scale,” says Computer Science and Engineering Prof. Lu Wang, the project’s lead investigator, “and little is known about how they interact with people’s inclination to have an impact and confirm their own values.”

This uncertainty stems from the problem’s scope: understanding the narratives used in news media, for example, and how they affect millions of unique individuals involves countless variables. In particular, Wang cited three major challenges that will need to be addressed in order to gain useful insights.

First, extracting narratives from source material requires a computational model that they can match to, and therefore a concrete understanding of what a narrative even is.

“It is critical to depict narratives with modularized representations,” Wang says, “to facilitate deeper understandings of narrative components and their effects on narrative influence.”

Next, she says, is that we’re missing knowledge of narrative shaping operations. The crafting of a narrative, whether it’s in the news or an advertising campaign, usually involves group collaborations that are lost in the final product. Lacking this “source material,” so to speak, makes the original intentions of the narrative hard to identify, and the real effects are typically varied and complex.

Finally, the ability to measure narrative influence at different scales is lacking. According to Wang, prior work in this area focuses mainly on the scale of individuals or of entire social media websites, but there’s little literature on smaller communities or how small-scale reactions build into large-scale reactions.

Wang’s goal is to conceptualize narratives using moral values that predispose humans to engage with and accept a narrative, and to detect and quantify the existence of narrative influence among individuals and communities. Her project will focus on natural language processing (NLP) and machine learning (ML) models to that end.

The project has three research thrusts:

  • Developing a generative model, based on social psychological theories, to represent narratives as multilevel framing products. This will be the first ever attempt to use moral values to create representative models of narratives that can be solved computationally.
  • Modeling group-level narrative crafting operations as an argumentation process aiming to maximize influence. Wang’s team will then mine and identify the argumentative structure of multiple articles and posts, and build graph neural networks to aggregate these arguments and enable prediction of narrative shaping tactics.
  • Building an NLP processing model to detect the targeted stances in an article or post, which are the building blocks that enable researchers to quantify individual and community changes in opinion. Along with a study of narrative influence among different demographics, this will enable better understanding of how to counter adversarial information operations.

Ultimately, Wang says, these three components will give researchers a better understanding of how information and attitudes flow through media into the opinions of communities and individuals. 

“This project will offer a deeper understanding of narratives and create new methods to automatically extract them,” says Wang. “The proposed methods and metrics are generalizable to individuals and groups of diverse demographics and different sizes, paving new ways for taming the spread of disinformation at the early stage.”