LabelboxDecember 15, 2020

University of Exeter uses ML to reveal how the media impacts human behavior

Researchers at the University of Exeter have been conducting two studies utilizing machine learning to explore the effects of news and social media on public attitudes and behavior.

In one study, a team consisting of researchers from Exeter, Trinity College Dublin, and the Exeter Q-Step Centre looks at how news media have framed the COVID-19 pandemic in both text and images. The project compares competing frames in the United States and the United Kingdom, and includes several news outlets with a diverse range of political leanings. The researchers seek to understand how the media has impacted attitudes and behavior around the global pandemic. This project requires a combination of convolutional neural network architecture and transfer learning to perform multi-label image classification.

One example of an image used by a news media outlet when reporting on COVID-19.

A second project, in partnership with the Exeter Q-Step Centre and the United States Forest Service, explores the extent to which urban environmental organizations in New York engage with nature in their social media communications. This study looks at how much these organizations use the natural environment in their communications with supporters. For this project, researchers are performing a segmentation analysis based on the U-Net architecture to complete their study.

An image shared by an organization on social media, labeled based on the presence of nature.

Both of these projects required a large amount of high quality training data however, and the tools and resources originally available to the team fell short. The team then discovered Labelbox in their search for a better option, drawn by the possibility of managing their own labeling team on the platform while leveraging the Labelbox workforce when necessary.

"This is, hands down, the best part of my experience with Labelbox so far," says Travis Coan, a researcher at Exeter, referring to his experience with the Labelbox Workforce. The labelers "were highly engaged, and the quality of the annotations was amazing."

Labelbox enabled the researchers to quickly scale up their data collection and provided the foundation for image analysis in both projects.

The combination of high-quality training data and recent advances in computational methods allowed them to explore key questions in political communication at a scale that had not previously been possible. With the large, rich dataset made possible by Labelbox and its workforce, the researchers can now more accurately determine the influence of the media on individual behavior.

Discover how the Labelbox Workforce could help you with your ML project.