Keng-Chi Chang

Keng-Chi Chang

Hello. I am an Assistant Professor in Quantitative Social Science at Dartmouth College.

I use computational tools, including AI and Large Language Models, to study how technology and media influence politics. My work explores censorship circumvention, online image-sharing behavior, and how short-form video content are recommended by algorithms. I’m fascinated by the ways our digital world is reshaping democracy and the possibilities—and risks—it creates for our social life.

My dissertation research has been funded by NSF/APSA and the Rapoport Family Foundation. My recent works have been published in the Proceedings of the National Academy of Sciences, Journal of Communication, PLOS One, and Journal of Quantitative Description: Digital Media, as well as peer-reviewed Computer Science conferences.

Interests

  • Computational Social Science
  • Political Methodology
  • Digital Politics
  • Image/Video as Data

Education

  • PhD in Political Science with a Specialization in Computational Social Science

    University of California, San Diego

  • MS in Computer Science

    University of California, San Diego

  • BA in Economics

    National Taiwan University

Research

The Importance of Prompt Tuning for Automated Neuron Explanations

The Importance of Prompt Tuning for Automated Neuron Explanations

Justin Lee, Tuomas Oikarinen, Arjun Chatha, Keng-Chi Chang, Yilan Chen, Tsui-Wei (Lily) Weng
NeurIPS 2023 Workshop ATTRIB, Accepted
  • Study neuron-level explanability by prompting LLMs with activated patterns.
  • Evaluate different prompting methods with both machine & humans.
Abstract PDF DOI
Do Imageries Lend Credibility to News Articles?

Do Imageries Lend Credibility to News Articles?

Working Paper
  • Varies image treatment while holding fixed other aspects of news articles.
  • Overall presence of image does not always increase credibility perception.
  • Images with some identified latent treatments (such as photos from press conferences, comics, or visuals of male suits) can alter credibility perception.
Abstract PDF
Characterizing Image Sharing Behaviors in US Politically Engaged, Random, and Demographic Audience Segments

Characterizing Image Sharing Behaviors in US Politically Engaged, Random, and Demographic Audience Segments

PhoMemes Workshop of 2023 AAAI ICWSM
  • Collect images shared by political & random Twitter users.
  • Study how sharing different types of images are predictive of demographic attributes.
Abstract PDF Slides
Compensation and the Consolidation of Authoritarian Power: Evidence from China’s 2016 PLA Reform

Compensation and the Consolidation of Authoritarian Power: Evidence from China’s 2016 PLA Reform

Working Paper
  • Collect large-scale biographical panel data of PLA officers.
  • Estimate patterns for promotion within PLA before/after PLA reform.
  • Finds that Xi only started to promote followers after power consolidation.
Abstract PDF Slides
Mapping Visual Themes among Authentic and Coordinated Memes

Mapping Visual Themes among Authentic and Coordinated Memes

PhoMemes Workshop of 2022 AAAI ICWSM
  • Collect coordinated IRA memes from Twitter and authentic memes from Reddit.
  • Cluster memes to find visual themes using self-supervised transfer learning.
  • Coordinated and authentic memes share visual themes but with different emphasis.
Abstract PDF Poster Slides DOI
COVID-19 Increased Censorship Circumvention And Access To Sensitive Topics In China

COVID-19 Increased Censorship Circumvention And Access To Sensitive Topics In China

Proceedings of the National Academy of Sciences
  • Use geolocated Tweets from China during the COVID-19 crisis.
  • Show that crisis motivates citizens to seek out sensitive information.
  • Gateway to both current and historically sensitive content is not found in countries without extensive online censorship.
Abstract PDF Code Slides DOI
The Effect of Streaming Chat on Perceptions of Debates

The Effect of Streaming Chat on Perceptions of Debates

Journal of Communication
  • Asks whether social chats on livestreams affect debate viewing.
  • Assign subjects to watch debates on ABC, Facebook Live, and 538.
  • Democratic subjects assigned to the Facebook chat condition reported lower affect towards Democrats and a worse viewing experience.
  • The tone of candidate-directed comments also matter.
Abstract PDF DOI
Using Facebook Data to Predict the 2016 U.S. Presidential Election

Using Facebook Data to Predict the 2016 U.S. Presidential Election

PLOS One
  • Use 19 billion likes to measure dynamic ideological positions of users and fan pages.
  • Guess user’s geolocation by likes and measure state level support rates for candidates.
  • Assume that users would support candidates with closer ideology.
  • FB support rates predict election outcome well, and share similar trends with polls.
  • Polls systematically overestimate Clinton’s support in right-leaning states.
Abstract PDF Code Slides DOI
Connective Effervescence and Streaming Chat During Political Debates

Connective Effervescence and Streaming Chat During Political Debates

Journal of Quantitative Description: Digital Media
  • Collect streaming chats of US Presidential debates on Facebook Live.
  • Describe their dynamics, degrees of toxicity, and insults.
Abstract PDF Code DOI
Social Mobility in Ming China: Evidence from Twelve Thousand Chin-shih Data

Social Mobility in Ming China: Evidence from Twelve Thousand Chin-shih Data

Working Paper
  • Challenge the conventional wisdom that civil service exam was meritocratic.
  • Use detailed data of past 3 generations to estimate the effect of family background on tests.
  • Multi-staged examination structure provides chances to partially control for ability.
Abstract PDF Slides