性如白玉烧犹冷     文似朱弦叩愈深

Forge a heart like jade—unyielding through the fire;

Write like plucking strings—press hard and ring higher.

Selected Research Projects

Not Just Bad Apples: A Large Language Model Approach in Studying Police Violence Accusation Framing at US Human Rights Reports

Accepted at Journal of Peace Research

Abstract: Does the US government's tarnished police violence record shadow its international human rights monitoring and reporting practices? Existing literature rarely considers the impact of US domestic human rights practices when examining the political bias in the US State Department's human rights (USSD) reports. In this project, I argue that, for the domestic struggles on the police violence issues, the US government tends to take a partial standing on reporting the related violations of other governments. With a novel network approach for text representation based on pre-trained large language models, the results show police violence accusations in the US human rights reports framed in favor of countries closer to the US. This research contributes to the human rights communities by highlighting the international impact of US domestic police violence issues and the existence of framing bias in human rights reports. The proposed method also shows promising potential in text analysis tasks like topic modeling.

Abstract: How does the language used during street protests affect the campaign outcomes? Using English protest signs in non-English speaking countries shows protesters’ intentions to attract support from international and specific domestic audiences. Meanwhile, English signs could also lend the government excuses for repression. In this article, we build a novel theory to explain the conditional merits of the strategic utilization of English signs in street protests. Relying on state-of-the-art artificial intelligence tools, we detected the language used in the protest signs for the campaigns from 2009 to 2019 from online protest photos. The empirical results suggest two implications. First, the effectiveness of English signs depends on the targeted governments’ vulnerability to international and domestic pressures. Second, using English signs on street protests might backfire and hurt the campaign. This article links the linguistic landscape research with the ongoing literature on campaign outcomes. The findings uncover the subtle mechanisms of how street protest strategy choice impacts campaign outcomes and also show the promise of image data in political research.

Gendered Imagery at Work: The Gender Influence of Protest News Reporting Using Audio Data

Under Review

Abstract: Do TV news correspondents report on female protesters differently? Despite the importance of media for civil resistance, current literature does not provide a clear prediction on the impact of the protesters’ gender in protest reporting. Drawing from the gender stereotype literature, I propose a conditional theory suggesting that for feminine stereotypes attached to female protesters to elicit positive media reactions, they must be activated, with a violent protest context serving as a significant external stimulus cue. To test the implications of this theory, I constructed a novel protest news report dataset, consisting of over 1,400 hours of protest news reporting videos from CNN and Fox News, developed with the aid of advanced AI tools. Using vocal pitch to measure correspondents’ emotional reactions during protest reporting and an innovative three-layer Bayesian Hierarchical Model to handle the nested structure inside of the data, empirical results indicate that only in violent contexts does identifying female protesters lead to notable emotional engagement from correspondents. This study introduces stereotype activation theory to gender studies in protest research, contributing to evolving literature on the audience effect of protesters’ identity and offering valuable insights for practitioners and researchers on the gender dynamics in civil resistance. The application of audio data and the proposed dataset also expands current methodological approaches in this field.

Abstract: What factors influence media attention on human trafficking? Estimates of individuals directly impacted by human trafficking have more than quadrupled in the last decade, even while state and non-state attention to the issue has grown. In this project, we examine the spread of media attention concerning human trafficking over time, focusing on how human trafficking is framed across print and broadcast sources. Using a variety of big data techniques, we create a novel dataset of human trafficking over time. While global awareness of the issue has increased, our analysis reveals notable “blind spots” and suggests missing populations in mainstreamed human trafficking discussions. Our project has implications for how scholars and advocates should analyze future social movements in both online and offline spaces and connects human trafficking to the broader literature on contentious politics and transnational advocacy.

Selected Research in Progress

AI Agents for Voting Simulation 

In Progress

Abstract: How would AI vote, and could the Large Language Model be used to simulate the voting behaviors in the US Presidential Election? In this research, I designed a debating game allowing seven AI agents to choose the presidential candidate they support and debate with each other to help their candidate win. The observational results suggest that (1) only the AI agents self-identified as white males will vote for Trump and all others will vote for Harris, even the voter agent who does not have any specified gender and race profile; and (2) AI agents are more strategic voters than ideological voters and tend to change their voting choice when their candidate is losing. 

"As We Got Your Attention": Research on Protest Reporting and Media Attention

In Progress

Abstract: What impact the media attention on protest reporting on different countries? Existing literature suggests an interactive system between the media and protesters in shaping the media attention distribution in protest coverage across countries and the privileged status of media in this interaction. Nevertheless, current empirical studies relying on the number of reports to measure media attention in protest reporting largely ignore the role of news audiences. Relying on advanced computer-vision algorithms, I propose a novel measurement of the media attention in news reporting while taking the news audiences’ perspective into consideration. Using a new dataset containing all the protest news on the frontpages of eight international news media websites, I apply a systematic examination of the impact of protest characteristics and media feature variables in shaping the media attention on protest coverage. The results suggest that, even though the variables from both strands could effectively shape the media attention in protest reporting, the variables of media features have much higher substantive impacts than that of protest characteristics. Moreover, the media attention circle and the global trend of protest news reporting could effectively encourage more media attention to report protests in specific countries. Eventually, there exists a zero-sum competition among the protesters from different countries for more media attention in protest coverage, and the US protesters enjoy a great media privilege. For protesters, the findings of the paper suggest that timing plays a pivotal role in navigating the media attention for the protest promotion.  

Grant Project: EMAI 

Funded by the College of Arts and Sciences, University of Alabama at Birmingham

Role: Co-Investigator

In Progress

Project goals: (1) design, model, and test a new AI-based measurement to assess human emotions that can be used as an additional scientific, evidence- based data source for qualitative researchers and (2) train and finetune the new algorithm for increased accuracy and better performance using qualitative interview recordings.