Deeper engagement is linked to criticism of a changing heritage village

The more deeply visitors engaged with Xiaozhou Village, an artist-settled urban heritage site in Guangzhou, China, the more likely their online accounts were to express negative views of the place. In an observational analysis of 12,493 posts and comments, 10.7% of accounts from people merely passing through were negative, compared with 22.8% from those describing a dedicated visit and 28.6% from those reporting repeated visits or extended stays. The researchers report that this contrast remained after adjustment for post length, platform, activity type and the writer’s position in the village, although the intermediate category became less distinct once author position was considered. The findings show an association, not that deeper engagement caused criticism, and suggest that familiarity and attachment may make environmental deterioration more perceptually and emotionally salient rather than shielding people from it.

For the study, the team assembled publicly available Chinese-language posts and comments from 13 platforms mentioning Xiaozhou Village, an informal settlement known for its waterways, narrow lanes, Qing-era oyster-shell masonry, old banyan trees and close relationship with a nearby art academy. The initial retrieval produced 39,921 records, from which the researchers removed irrelevant, duplicated, empty and promotional material. Their final corpus contained 2,850 posts and 9,643 comments. Because the same place name also referred to other Chinese locations, a television drama and unrelated tourism material, every retrieved record was individually assessed for relevance. Publicly visible records were analyzed without recruiting participants or contacting their authors; identifiers and contact details were removed or masked.

To quantify expressed sentiment, the researchers first used a Chinese language model to annotate records with scoreable text on a scale from strongly negative to strongly positive, then fine-tuned a 102.3-million-parameter Chinese RoBERTa transformer as a three-class classifier. Of the 12,493 records in the corpus, 11,741 had text usable for annotation and modeling; 752 comments contained only emoji, stickers or other material that could not be scored and were excluded from the models. The classifier was trained on 8,218 records and tested on 1,762, achieving an accuracy of 0.803 and a macro F1 score of 0.792. A second independently developed model re-annotated a stratified sample of 498 posts, producing a Pearson correlation of 0.914 and substantial categorical agreement.

The researchers also coded three levels of engagement: passing or informational mention, one purposive visit, and repeated visits or extended dwelling. Independent re-coding of 499 posts gave 79.0% exact agreement for engagement depth, with Cohen’s kappa of 0.695 and a quadratic-weighted kappa of 0.724. The analytical model treated 18 binary lexicon features, platform indicators and post length as predictors of a continuous sentiment score. These features covered vernacular architecture, ecological elements, artistic practices, atmospheric qualities, consumption and specific problems such as crowding, commercialization, access, dilapidation and disappointment. Gradient boosting was selected after comparison with ordinary least squares and random forest models, with a ten-fold cross-validated R² of 0.137, root mean squared error of 0.635 and mean absolute error of 0.495.

That explanatory performance was modest: the 18 textual cues accounted for only a limited share of the variation in sentiment and were not intended to capture the full determinants of visitors’ experiences. SHAP, or Shapley additive explanations, was used to estimate each feature’s marginal contribution to model predictions rather than its causal effect in the physical environment. The strongest positive model contributions came from narratives of discovering a “hidden gem,” with a mean SHAP value of 0.221 and 95% confidence interval of 0.210 to 0.232; food and consumption contributed 0.205, and descriptions of quietness or slow pace contributed 0.157. Ancient banyans contributed 0.152, oyster-shell walls 0.150, photography 0.097 and sketching 0.087.

The largest negative contributions were expressions of disappointment, at −0.766 with a 95% confidence interval of −0.810 to −0.721, and references to dilapidation and litter, at −0.583 with a 95% confidence interval of −0.618 to −0.550. Crowding, cost and ticketing, access and parking, and over-commercialization also received negative contributions, although generally smaller ones. Because disappointment appeared in 3.0% of posts and dilapidation in 3.9%, the authors argued that overall positivity persisted more because strong negative themes were uncommon than because positive cues consistently outweighed them. These attribution values characterize the fitted model; they do not establish that particular features cause favorable or unfavorable evaluations.

Activity type alone did not explain the pattern. After adjustment for post length and platform, sketching had no detectable association with whether environmental problems were mentioned, with an odds ratio of 1.12 and p-value of 0.795, while photography had a similar null association with problem mentions at 1.05 and p=0.863. Photography, however, was associated with markedly less negative expression overall, with an odds ratio of 0.43 and p=0.0006, whereas sketching showed no detectable difference at 0.75 and p=0.381. A corrected lexicon was critical to that conclusion. An earlier feature combined the act of drawing or painting with terms for studios, academies, examination ateliers and graffiti, producing an apparent association between sketching and negativity. Separating artistic practice from the institutions surrounding it removed the result; posts mentioning institutional art terms were more negative, even though no such difference appeared among people actually describing painting.

Engagement depth remained the more consequential variable. Among 2,686 posts for which depth could be established, negative sentiment rose across the three levels on the annotation measure. Models controlling for post length and platform produced odds ratios of 2.12 for the intermediate category and 3.25 for repeated visitors or residents relative to passers-by, while post length had no independent association with negativity at 0.97 and p=0.523. Adding activity variables left the pattern largely unchanged. A sensitivity analysis excluding posts that contained either painting-practice or art-institution terms also retained the gradient, with negative shares of 11.0%, 23.7% and 27.5%. When author position was added, however, the intermediate contrast weakened and the odds ratio for the most deeply involved group fell from 3.29 to 2.14. The most defensible conclusion is therefore a robust contrast between sustained involvement and passing contact, rather than a perfectly monotonic progression through all three categories.

Who wrote also mattered. Among 1,795 posts with identifiable author positions, local residents were the most critical: 33.3% of their 72 posts were negative. General visitors had a negative share of 22.8%, while media or promotional accounts had 7.5% and commercial accounts 1.0%. Established artists were an exception, with 13.0% negative posts among 46 posts attributed to that group, although their relatively favorable result did not persist after adjustment. The authors say several explanations remain possible: established artists may hold commercial premises and therefore have incentives to protect the village’s public image, or artists most dissatisfied with local conditions may already have left. The commercial concentration of positive commentary also means that aggregate platform sentiment cannot be treated as a simple proxy for the views of residents or visitors, many of whom never post.

The result challenges theories predicting that stronger engagement should make a place seem better. Processing-fluency and effort-justification accounts suggest that repeated exposure or investment can encourage favorable evaluation, while recreation specialization and solastalgia predict more exacting judgments from specialists and people attached to places undergoing change. Xiaozhou’s pattern aligned with the latter: sustained involvement supplied both a standard against which the village could be judged and a stake in whether it met that standard. Yet naming specific conditions did not rise monotonically with depth; deeply involved authors often discussed deterioration generally rather than listing individual problems, so the central outcome was negative expression rather than the frequency of environmental terms. The study does not establish whether such critical expression reflects greater attention, greater attachment, selection of more concerned writers or some combination of these factors.

Several limitations prevent a causal or population-wide interpretation. The study was cross-sectional, relied on what people chose to publish and inferred engagement from text rather than surveying participants. One author could contribute multiple posts, but stable identifiers were unavailable, while comments were nested within posts; the analysis did not correct for either clustering, meaning its confidence intervals may be too narrow. The corpus spans 2008 to 2026, with coverage becoming much denser after 2018, and Xiaozhou was treated as a theoretically informative single case rather than a statistically representative sample. Reverse causation also remains possible: people already predisposed to notice decline may be more likely to become deeply involved. The evidence therefore supports an association between sustained engagement and critical expression, not proof that time spent in Xiaozhou produces negativity. It also suggests that aggregate visitor satisfaction may remain high even when the people most dependent on a place’s environmental conditions express more concern.

Subject of Research: How engagement depth is associated with negative expressions about environmental decline in Xiaozhou Village, Guangzhou.

Article Title: Attentional reorganization in heritage environments: depth of engagement and the expression of environmental decline

Article References: Li, S., Xia, Q., Huang, S., Huang, S., Lv, Y., Jiang, Y., & Chen, X. (2026). Attentional reorganization in heritage environments: depth of engagement and the expression of environmental decline. Frontiers in Psychology, 17, Article 1800968. https://doi.org/10.3389/fpsyg.2026.1800968

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1800968

Keywords: environmental psychology, heritage tourism, place attachment, Xiaozhou Village, social media analysis, explainable AI, sentiment analysis, solastalgia, urban villages, environmental perception, Attentional, reorganization