Machine learning maps how 197 million Chinese government texts signal reputation

In the largest computational study of its kind, researchers have classified roughly 197 million official web articles published by Chinese local governments to reveal how administrative bodies at different tiers of the state broadcast their credibility to the public. The analysis, published in Frontiers in Political Science, shows that provincial, prefecture-level municipal, and county-level governments share a common architecture of reputational signaling, yet each level tilts its communication toward a distinct dimension of bureaucratic reputation. Municipal governments foreground concrete achievements, county governments emphasize moral responsibility and public care, and provincial governments stress procedural legitimacy, while all three levels give comparatively little attention to technical expertise.

The study, conducted by Youkui Wang of Beijing University of Posts and Telecommunications, Xu Geng of Shenzhen University, Sicheng Chen of Nanjing University, and Qingguo Meng of Tsinghua University, introduces the concept of a reputational signal portfolio: the relative allocation of an organization’s communication across four claim types drawn from the multidimensional theory of bureaucratic reputation. Performative signals communicate the capacity to achieve goals, including policy accomplishments, administrative efficiency, and successful implementation. Moral signals emphasize public responsibility, responsiveness, and alignment with collective values. Procedural signals highlight compliance with rules, transparency, consultation, and legitimate decision-making. Technical signals convey expertise, specialized knowledge, and problem-solving ability. Because organizations face multiple audiences and limited communicative attention, the prominence of any one signal is meaningful largely in relation to the others in the portfolio.

The empirical foundation is unusually broad. The researchers compiled official web articles from 31 provincial-level governments, 281 prefecture-level municipal governments, and 2,020 county-level governments over a continuous 60-month window from January 2015 to December 2019. After removing duplicates, invalid pages, and non-textual materials, the corpus contained approximately 197 million records. Government websites in China are not voluntary publicity channels; national regulations on information disclosure and the 2017 Guidelines for the Development of Government Websites formalize information publication, policy interpretation, online services, and public interaction as core functions, embedding website content in administrative procedures and accountability arrangements.

Because manual coding at this scale is impossible, the team combined theory-guided annotation with supervised machine learning. A random sample of 10,000 article titles was drawn from the preprocessed corpus, and the title, rather than the full text, served as the unit of classification, since titles on Chinese government portals are concise and typically state the central event, policy action, or reported outcome directly. Three researchers independently assigned each title to one of four mutually exclusive categories, aided by a keyword dictionary that functioned as a coding aid rather than a deterministic rule. In the initial round, all three coders agreed on 62.53 percent of titles, and at least two agreed on an additional 28.86 percent, meaning 91.39 percent received a majority classification before discussion. All disputed cases were resolved by consensus, yielding a final labeled sample of 3,138 performative, 3,126 moral, 2,315 procedural, and 1,421 technical titles.

The team compared three algorithms implemented in Python with scikit-learn: Naive Bayes, Random Forest, and Gradient Boosting Decision Trees. Gradient boosting performed best and became the final classifier, achieving an overall held-out test-set accuracy of 0.7807, with F1 scores above 0.70 for the performative, moral, and procedural categories and above 0.60 for the technical category. Once validated, the classifier was applied to every retained title in the corpus, and the results were aggregated by government and month. For each government-month observation, the emphasis on a dimension was measured as the share of titles classified into that category out of all four categories, so the four indicators sum to one and capture relative composition rather than absolute communication volume.

The aggregate findings show a stable but uneven portfolio. Performative reputation accounts for 30.79 percent of identified signals and moral reputation for 31.92 percent, together dominating official communication, while procedural reputation represents 21.34 percent and technical reputation 15.95 percent, the least emphasized dimension. The authors stress that these measures capture communicative visibility, not organizational capacity or audience evaluation: a large share of performative signals does not demonstrate strong performance, and limited technical signaling does not imply weak professional expertise, since technical work may be central to operations but harder to communicate to broad audiences than achievements or welfare commitments.

Hierarchical differences emerged clearly within this shared structure. Municipal governments devoted a significantly larger share of their portfolios to performative signals than provincial governments, 32.29 percent versus 30.05 percent, consistent with their central role in policy implementation, economic development, infrastructure, and urban governance, where they are frequently evaluated through targets and development indicators. Provincial governments placed greater emphasis on procedural signals, 23.31 percent versus 21.61 percent for municipalities and 21.32 percent for counties, consistent with their responsibilities for regional coordination, planning, rule formulation, and supervision of subordinate governments. County-level governments devoted the largest share to moral signals, 32.23 percent, compared with 29.88 percent for provincial and 29.75 percent for municipal governments, a pattern the authors link to their direct involvement in education, healthcare, social assistance, rural development, and everyday services that bring them into close contact with citizens. Pairwise comparisons across all four dimensions were statistically significant at the 1 percent level, although the authors caution that small differences, particularly for technical reputation, should not be equated with substantive magnitude.

The portfolios also proved temporally dynamic. Descriptive regression models using year, month, and administrative level as predictors, based on 139,920 government-month observations across 2,332 governments, indicated an overall positive association between year and the relative shares of performative and procedural signals and a negative association with moral and technical signals, though year-specific comparisons showed these changes were not uniformly monotonic, with procedural signals showing the clearest longer-term increase relative to 2015. Within years, the four dimensions followed distinct seasonal rhythms: performative signaling peaked in the second quarter before declining in the second half of the year, moral signaling was strongest in the first quarter, procedural signaling rose progressively from the second to the fourth quarter and was especially elevated in November and December, and technical signaling was more visible in the second and third quarters. These patterns are consistent with recurring administrative activities such as annual policy deployment, implementation, inspections, budgeting, information disclosure, and year-end reporting, and they unfolded against an evolving institutional environment that included the first nationwide government-website inspection launched in 2015 and the 2017 standardization guidelines.

The authors are explicit about what the design cannot establish. Because the four portfolio shares sum to one, the dimension-specific coefficients describe changes in relative prominence rather than independent changes in absolute volume, and the separate models do not explicitly account for this compositional dependence. Repeated monthly observations within governments may exhibit within-government dependence not fully modeled in the reported specifications, so the regression results should be read as descriptive evidence of systematic associations rather than causal estimates. The observational design also cannot directly identify the mechanisms proposed to explain hierarchical differences, namely functional differentiation, accountability relationships, and audience proximity, which the authors present as theoretically grounded interpretations rather than established causal pathways. Nor can official texts reveal organizational intentions: observed signals may reflect strategic reputation management, legal disclosure requirements, policy campaigns, or routine reporting, and the study does not attribute a common intention to every observed change.

Conceptually, the study shifts attention from the reception of reputation, which has dominated organizational research, to its supply side: the informational environment that public organizations themselves construct before audiences form evaluations. By demonstrating that reputational signaling varies systematically across administrative levels and over time while retaining a stable underlying architecture, the analysis suggests that reputation management in multilevel governance is neither a fixed organizational trait nor a continuously redesigned strategy, but an ongoing administrative practice characterized by substantial stability alongside bounded, patterned adjustment. The authors note that the machine-learning approach complements rather than replaces surveys, experiments, and qualitative methods, which remain essential for examining audience perceptions and organizational intentions, and they propose linking these population-level signal measures to surveys, media coverage, and online public responses to test whether different portfolios are associated with trust, cooperation, and perceived government competence in China and other political systems.

Subject of Research: Reputational signaling portfolios in Chinese multilevel government communication measured by supervised machine learning

Article Title: Governing reputation at scale: machine-learning evidence from 197 million official texts across China’s multilevel government

Article References: Wang, Y., Geng, X., Chen, S., & Meng, Q. (2026). Governing reputation at scale: machine-learning evidence from 197 million official texts across China’s multilevel government. Frontiers in Political Science, 8, 1951998. https://doi.org/10.3389/fpos.2026.1951998

Image Credits: AI Generated

DOI: 10.3389/fpos.2026.1951998

Keywords: organizational reputation, bureaucratic reputation, machine learning, computational text analysis, administrative hierarchy, China, multilevel governance, government websites, reputational signal portfolio, supervised classification, public administration, longitudinal text analysis