My published research examines how digital technologies, social relationships, and institutional contexts shape people’s experiences and opportunities. I combine computational social science and quantitative methods to study teachers’ use of social media, interaction in digital environments, social networks, and the design of credible social research.
Publications are grouped by theme below. Each entry includes an abstract or a labeled abstract summary and a link to the published work. For a full publication list, see my Google Scholar profile and curriculum vitae.
Social Interaction, Networks, and Well-being
These studies examine how the content and structure of social relationships relate to participation, support, and well-being in learning environments and everyday life.
The ties that bind: migrant worker social networks, mental health, loneliness and self-rated health
Timothy S. Sumerlin, Zixi Chen, Edward W. W. Chan, Carl Latkin, and Brian J. Hall. 2026. Epidemiology and Psychiatric Sciences, 35, e58.
Abstract
Aims. Migrant domestic workers (MDWs) are separated from family and at-home support networks while working abroad. This may affect their mental health due to stressors, social isolation and loneliness. Evidence shows mixed results on whether social capital and social support enhance MDWs’ mental health, warranting further investigation. This study assessed how positive and negative dimensions of social network capital are… Show moreShow less
Aims. Migrant domestic workers (MDWs) are separated from family and at-home support networks while working abroad. This may affect their mental health due to stressors, social isolation and loneliness. Evidence shows mixed results on whether social capital and social support enhance MDWs’ mental health, warranting further investigation. This study assessed how positive and negative dimensions of social network capital are associated with mental health, loneliness and self-rated health among MDWs in Macao SAR, China.
Methods. Using respondent-driven sampling, 1,681 MDWs (Filipino n = 1329; Indonesian n = 352) were enrolled. Egocentric name-generating questions captured network size, emotional closeness, social support helpfulness, reciprocity burden and financial support access. Multivariable linear regression examined associations stratified by nationality, and subgroup analysis by network type were examined. To address multiple comparisons across five network capital measures and five outcomes, p-values were adjusted using the Benjamini–Hochberg (BH) false discovery rate procedure.
Results. Network capital characteristics differed across network types. The mean (standard deviation) overall network size was 6.00 (3.77) for Filipino MDWs and 3.69 (2.52) for Indonesian MDWs. Following BH adjustment, among Filipino MDWs, greater social support helpfulness was associated with lower loneliness (Standardized β [Std β] = −0.15, Standard error [SE] = 0.02, p < 0.001) and better self-rated health [Std β = 0.15, SE = 0.01, p < 0.001]), while higher reciprocity burden was consistently associated with poorer outcomes (e.g., post-traumatic stress disorder [PTSD]: Std β = 0.21, SE = 0.63, p < 0.001). Network size was associated with higher mental health symptom scores (Depression: Std β = 0.08, SE = 0.04, p = 0.004; Anxiety: Std β = 0.08, SE = 0.04, p = 0.003; PTSD: Std β = 0.08, SE = 0.11, p = 0.003), but less loneliness (Std β = −0.07, SE = 0.01, p = 0.007). Among Indonesian MDWs, reciprocity burden and financial support access showed associations primarily with depression (reciprocity burden: Std β = 0.44, SE = 0.60, p < 0.001; financial support: Std β = 0.22, SE = 0.97, p = 0.007) and anxiety (reciprocity burden: Std β = 0.32, SE = 0.60, p < 0.001; financial support: Std β = 0.23, SE = 0.97, p = 0.008) within roommate networks. Most standardized effect sizes were modest (Std β < 0.25).
Conclusions. Network capital is associated with mental health, loneliness and self-rated health among MDWs, but patterns differ by nationality and network type. Culturally responsive and context-specific strategies that strengthen supportive ties and reduce network-related burdens may improve MDWs’ health.
Why do students reply? Uncovering the socio-semantic entanglement in web annotation activities
Bodong Chen and Zixi Chen. 2024; first published online in 2023. Education and Information Technologies, 29(8), 9747–9771.
Abstract
Web annotation environments are widely used in education based on the premise that student interaction in these environments benefits individual and group learning. However, there is little research on factors driving student interaction in web annotation activities. In this study we asked: What dynamics could explain social interaction among students in web annotation activities in college classrooms? Recognizing the mediated… Show moreShow less
Web annotation environments are widely used in education based on the premise that student interaction in these environments benefits individual and group learning. However, there is little research on factors driving student interaction in web annotation activities. In this study we asked: What dynamics could explain social interaction among students in web annotation activities in college classrooms? Recognizing the mediated nature of social interaction in web annotation, we hypothesized that student interaction is driven by multiple factors including previous relational events and semantic features of annotation content. Following a novel network analysis method named relational event modeling, we analyzed a rich dataset from four online classes. Results indicated annotation popularity was initially predictive of student replies, meaning popular annotation threads were more likely to attract new replies. However, this effect nearly diminished when adding thread-level semantic cohesion to the model, indicating an significant role played by semantic cohesion in attracting new responses. This paper makes important progress towards modeling social interaction in digital environments as a dynamic, mediated phenomenon. This study contributes empirical insights into web annotation and calls for future work to investigate social interaction as a dynamic network phenomenon.
AI and Language
These publications examine AI in social communication and investigate how multilingual models represent language and meaning.
When AI Joins the Thread: A Computational Analysis of Gendered Human-AI Interactions on Weibo
Chenxi Li, Zeqiang Wang, Yujia Wang, Jon Johnson, Suparna De, and Zixi Chen. Published online in 2026. In Social Computing: ICSC 2025, Revised Selected Papers, 137–150.
Abstract
This study investigates how a Weibo-embedded artificial intelligence (AI) chatbot, CommentR, functions within a gendered communicative environment by systematically mapping the conversational contexts and linguistic patterns through which gendered dynamics manifest in human-AI interaction. Drawing on a large-scale dataset of interactions between Weibo users and CommentR, the study identifies the dominant thematic structures and user intentions that characterize engagement with… Show moreShow less
This study investigates how a Weibo-embedded artificial intelligence (AI) chatbot, CommentR, functions within a gendered communicative environment by systematically mapping the conversational contexts and linguistic patterns through which gendered dynamics manifest in human-AI interaction. Drawing on a large-scale dataset of interactions between Weibo users and CommentR, the study identifies the dominant thematic structures and user intentions that characterize engagement with the chatbot, while quantifying stylistic variations in its replies across user genders. Through the integration of computational methods and large language model (LLM)-based annotation, the analysis reveals how subtle linguistic features vary across user gender and interactional context. Ultimately, the findings provide an empirical basis for understanding how everyday exchanges with AI systems both reproduce and reshape gendered communication norms within China’s digital public sphere.
Decomposing Multilingual Representations: How Scale, Architecture, and Data Shape Functional Specialization
Zeqiang Wang, Xinyue Wu, Chenxi Li, Yuqi Wang, Zixi Chen, Jon Johnson, and Suparna De. 2026. Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 17487–17491.
Abstract
How Large Multilingual Models (LMMs) separate linguistic form from semantic content is largely opaque. This paper introduces a framework to dissect their internal representations, revealing a phenomenon we term Functional Specialization: the emergence of distinct neural circuits for language-specific form versus language-agnostic semantics. Through extensive experiments on the E5 and Qwen series, we show this specialization is governed by three… Show moreShow less
How Large Multilingual Models (LMMs) separate linguistic form from semantic content is largely opaque. This paper introduces a framework to dissect their internal representations, revealing a phenomenon we term Functional Specialization: the emergence of distinct neural circuits for language-specific form versus language-agnostic semantics. Through extensive experiments on the E5 and Qwen series, we show this specialization is governed by three factors. Architecture dictates the core strategy: encoders adopt specialization breadth (many features in a staged workflow), while decoders pursue specialization depth (few, high-purity features). Scale primarily drives neural efficiency, enabling robust separation with fewer circuits. Finally, high-clarity data acts as a catalyst, inducing sophisticated mechanisms even in smaller models. Our findings chart a path toward more controllable and interpretable multilingual AI.
Research Design and Generalizability
How far can a finding travel beyond the sample in which it was observed? This work develops tools for assessing the robustness of conclusions about a target population.
The Probability of a Robust Inference for External Validity: A Probabilistic Generalizability Index for Randomized Experiments
Tenglong Li, Mingming Chen, Yi Lin, and Zixi Chen. 2026; first published online in 2025. The Journal of Experimental Education, 94(4), 1151–1164.
Abstract
External validity is often questionable in empirical research, especially in randomized experiments due to the tradeoff between internal validity and external validity. To quantify the robustness of external validity, one should first conceptualize the part of the target population that cannot be represented by the observed sample (i.e., the unobserved part) and thus the unobserved sample which is thought of… Show moreShow less
External validity is often questionable in empirical research, especially in randomized experiments due to the tradeoff between internal validity and external validity. To quantify the robustness of external validity, one should first conceptualize the part of the target population that cannot be represented by the observed sample (i.e., the unobserved part) and thus the unobserved sample which is thought of as a random sample drawn from the unobserved part of the target population. In this article, we define the probability of a robust inference for external validity, i.e., the PEV, as the probability of rejecting the null hypothesis again based on both the observed and unobserved samples, given a significant result based on the observed sample. Drawing on the relationship between the unobserved sample and the PEV, we propose a six-step procedure for evaluating external validity and illustrate this procedure with an empirical example. We show that the PEV can be interpreted as the statistical power and our analysis of external validity can be translated as power analysis across all plausible forms of the unobserved sample.
Organizations and Institutions
These studies examine organizational arrangements and people’s experiences within them, from university research laboratories to workplace conditions expressed in online reviews.
Listening to your employees: analyzing opinions from online reviews of hotel companies
Xiaolin (Crystal) Shi and Zixi Chen. 2021. International Journal of Contemporary Hospitality Management, 33(6), 2091–2116.
Abstract summary
This study uses 78,535 employee reviews of 29 hotel companies posted on Indeed between 2011 and 2019 to investigate workplace satisfaction. Structural topic modeling and sentiment analysis identify 20 topics and compare employees’ accounts by hotel category and employment status. Workplace bullying and dirty work, organizational support, and career advancement were the most prevalent topics. Reviews of premium hotels were… Show moreShow less
This study uses 78,535 employee reviews of 29 hotel companies posted on Indeed between 2011 and 2019 to investigate workplace satisfaction. Structural topic modeling and sentiment analysis identify 20 topics and compare employees’ accounts by hotel category and employment status. Workplace bullying and dirty work, organizational support, and career advancement were the most prevalent topics. Reviews of premium hotels were more positive overall, while sentiments about particular topics varied between current and former employees and between premium and economy hotels. The study shows how online employee accounts can complement surveys when examining workplace experiences and turnover.