The researchers assembled an unusually large corpus: 64,349 publicly visible top-level comments posted in response to the donated-artifact controversy. Rather than treating the comment stream as an undifferentiated mass, they organized it into three event-anchored source-content groups, each tied to a distinct moment in the unfolding crisis. This design choice matters. Crisis discourse is not static; the public’s demands in the hours after an initial institutional response can differ sharply from those expressed once investigations are underway. By anchoring comment groups to specific events, the team could track how the texture of accountability talk shifted over the life of the controversy, rather than averaging it into a single blurred snapshot.

The computational backbone of the study is BERTopic, a transformer-based topic modeling technique that has become a workhorse of large-scale discourse analysis. Unlike classical topic models such as latent Dirichlet allocation, BERTopic leverages contextual embeddings from pretrained language models, clustering documents by the semantic similarity of their vector representations and then extracting interpretable keywords for each cluster. Applied to the Nanjing Museum corpus, the method identified 49 substantive topics that together covered 42,126 comments—a substantial and coherent subset of the full data. The remaining comments fell outside the substantive clusters, a reminder that even the best topic models leave residual noise at the messy edges of real-world social media text.

Topic models produce clusters, but clusters are not theory. To move from 49 granular topics to interpretable findings, the researchers applied a higher-order thematic framework, validated independently by two researchers, that mapped the topics onto six substantive groupings. These groupings were then aligned with a concept the authors call distributed accountability—a function-based framework for understanding how public demands for sanction, explanation, and relational repair are organized across platform environments. In this scheme, accountability is not a single demand but a bundle of distinct functions: punitive accountability seeks punishment and blame; knowledge-based accountability demands investigation, evidence, and explanation; ethical accountability interrogates the moral norms governing behavior, in this case the ethics of donating and accepting cultural artifacts.

The headline quantitative result is that ethical accountability accounted for the largest share of classified comments under the primary coding specification. In other words, when Chinese social media users confronted the museum controversy, the dominant mode of judgment was moral rather than punitive or investigative: commenters were most engaged by questions of what ethical standards should govern donations to a major public institution. Yet the authors are refreshingly candid about the fragility of that ranking. The absolute ordering of accountability functions proved sensitive to how the largest disputed topic was treated—a methodological honesty that distinguishes careful computational social science from the overconfident claims that sometimes plague the field.

More robust were the temporal patterns visible across the event-anchored groups. Donation ethics was relatively more prominent in discussion attached to the institution’s initial response, when the moral contours of the controversy were freshest and most contested. As the crisis matured, investigation-oriented themes and personalized blame grew relatively more prominent, suggesting that public attention migrated from abstract ethical principle toward concrete demands for process and identifiable responsibility. This trajectory echoes a well-documented pattern in crisis communication research: initial moral shock gives way, over time, to procedural scrutiny and the search for culprits. What the study adds is evidence that this migration can be measured at scale, comment by comment, across the full arc of a cultural heritage crisis.

The cross-platform findings are the study’s most conceptually interesting contribution, and also its most carefully hedged. Douyin, the short-video platform, showed a relatively stronger punitive profile: its comment sections leaned toward demands for sanction and blame. Bilibili, the community built around longer-form video and known for its technically literate user base, showed a relatively stronger knowledge-based profile, with more emphasis on investigation and explanation. Rednote occupied its own discursive position within this triad. Crucially, these platform differences were stable across alternative coding specifications, which strengthens confidence that they reflect genuine differences in how communities organize accountability talk rather than artifacts of one particular analytical choice.

But the authors resist the seductive conclusion that platforms create entirely distinct publics. The overall association between platform and accountability function, while detectable, remained small. As the study puts it, platform differences concern the relative organization of accountability discourse, not the presence or absence of particular accountability functions. Every platform hosted punitive, knowledge-based, and ethical demands; what varied was their proportionate weight. This finding pushes back against a popular tendency to treat platforms as hermetically sealed judgment machines—Douyin as the mob, Bilibili as the analysts—and instead supports a picture of functionally differentiated accountability demands coexisting within a shared institutional crisis, redistributed but not partitioned by platform architecture.

The theoretical payoff is twofold. For scholars of platformized public judgment, the study demonstrates that accountability can be decomposed into measurable functions and tracked across platform boundaries, offering a template that future researchers can apply to other crises. For practitioners of institutional crisis governance, the implications are practical: a museum or heritage body navigating a scandal cannot craft a single message and broadcast it uniformly. The public’s demand for explanation will be amplified in some spaces, its demand for sanction in others, and its demand for relational repair—acknowledgment, apology, restored trust—everywhere. Governance strategies that ignore this distributed structure risk answering the wrong question in the wrong venue at the wrong moment.

The Nanjing Museum controversy will eventually fade from feeds, but the methodological lesson endures. By combining transformer-based topic modeling with a theoretically grounded accountability framework, event-anchored sampling, and explicit sensitivity analysis, Li, Ma, and Lu have shown how the anatomy of public judgment can be dissected without flattening it. Cultural heritage institutions, which increasingly find their crises adjudicated in real time by millions of strangers across incompatible platforms, now have a sharper map of the terrain they must navigate: not one court of public opinion, but many overlapping ones, each weighing punishment, proof, and principle in its own proportions.

Subject of Research: Cross-platform topic modeling of public accountability discourse during the Nanjing Museum donated-artifact controversy

Article Title: Cross-platform public judgment in a cultural heritage crisis: A topic modeling study of the Nanjing Museum donated-artifact controversy

Article References: Li, W., Ma, M., & Lu, J. (2026). Cross-platform public judgment in a cultural heritage crisis: A topic modeling study of the Nanjing Museum donated-artifact controversy. PLOS One, 21(10), e0360658. https://doi.org/10.1371/journal.pone.0360658

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0360658

Keywords: Nanjing Museum, donated artifacts, cultural heritage crisis, BERTopic, topic modeling, distributed accountability, Douyin, Bilibili, Rednote, crisis communication, platform studies, public judgment