Monthly updates from the Policy Change Index projects.
Mercatus Center at George Mason University This Week
 


Dear Human Readers,
 
We are excited to announce the launch of our latest machine learning algorithm, the Policy Change Index for Outbreak (PCI-Outbreak)!
 
It’s widely suspected that China’s official COVID numbers understate the scale of the outbreak. But by how much? The PCI-Outbreak tackles this question by “reading” the tea leaves in China’s flagship newspaper, the People’s Daily. The project was featured in yesterday’s China Watcher newsletter curated by Politico’s David Wertime.
 

PCI-Outbreak

Words speak louder than (fake) numbers. Here’s the idea: It’s easy to outright release false statistics, but it’s harder to conceal the truth from the public when the government has to talk about the crisis at length in national media. To wit, when the Chinese authorities announced the lockdown of Wuhan, a city with 11 million people, only 500-odd cases had been confirmed nation-wide.
 
The curve was “flattened,” but rather slowly. The chart below shows the word-based PCI-Outbreak indicator, on a scale of 0 to 1, in comparison with the official statistics. The numbers say the outbreak hit plateau in February, which checks out. But our algorithm suggests the severity did not decline nearly as quickly as what the numbers let on—hence the big gap between the two curves. Moreover, the index spiked up again in June as Beijing saw another outbreak of cases.

Figure 1: PCI-Outbreak and official COVID cases in China (Jan 21 - Jul 9)

Note: The PCI-Outbreak learns the change in language in China’s media coverage of the 2003 SARS outbreak and then assesses how serious-sounding the coverage of COVID-19 is.

Learning from the past... Our program is able to see that because it has learned from the 2003 SARS outbreak to understand the wax and wane of language as that epidemic went through its cycle. With the past as the benchmark, the COVID language becomes indicative of how severe the current outbreak is.
 
… with BERT! This approach may sound familiar to our regular readers: the PCI-Crackdown also looks at the 2019-20 Hong Kong protests through the lens of the 1989 Tiananmen Square protests. A major difference: This time, we have adopted the Bidirectional Encoder Representations from Transformers (BERT), famously developed by Google AI in 2018. Check out the PCI-Outbreak’s source code here.

PCI-Outbreak is developed by Weifeng Zhong, Julian TszKin Chan, Kwan-Yuet “Stephen” Ho, Kit Lee, and Kawai Leung.
 

PCI Resources

The open-sourced PCI projects are meant to crack a window to otherwise opaque political systems, so everyone can look inside—for free. You can find out more about the projects on the PCI website. Don’t hesitate to reach out!

 

Edited by Weifeng Zhong and Julian TszKin Chan

 

If you were forwarded this email and like what you read, please sign up here to stay in the loop. If you’re already a loyal reader, please help us grow by forwarding this to a friend or colleague. 
 
Subscribe →
Copyright © 2020, All rights reserved.

Mercatus Center at George Mason University, 3434 Washington Blvd, Arlington VA, 22201