The research team, led by Wenlei Niu and Feng Liang of Wuxi University of Technology together with Qiongxuan Zeng of Chenzhou Vocational Technical College and Yingpeng Yu, examined reviews and photographs from Yuelu, Bailudong, Songyang, and Yingtian academies. Songyang Academy forms part of the UNESCO World Heritage property Historic Monuments of Dengfeng, while Bailudong Academy sits within the broader Mount Lushan cultural landscape. The four sites share an educational-heritage function but differ in historical development, physical setting, and contemporary tourism context, so the researchers included academy as a site-level covariate in their multivariable analyses. Reviews and photographs were retrieved on 28 February 2026, with recorded review dates ranging from March 2016 to February 2026.

At the heart of the study is a conceptual distinction that the authors argue is often blurred in multimodal heritage research. A visitor’s uploaded photograph records what that person chose to show; a written review records what they chose to discuss; an explicit judgment about a particular environmental quality, such as tranquility or artificiality, is not the same as the overall evaluative tone of the review; and none of these is equivalent to a platform star rating. The team organized these into five observational layers: visitor-selected visual content, heritage-related discourse, explicit environmental-quality evaluation, textual valence, and platform star ratings.

To build the analysis, the researchers classified each sentence into one of seven topics using multilingual sentence embeddings with BERTopic and KMeans clustering, covering historical-cultural discourse, scholarly atmosphere, architectural heritage and spatial form, interpretation and knowledge, environmental landscape and leisure experience, service and access, and on-site process. A two-author validation of 300 stratified sentences yielded a Cohen’s kappa of 0.687, with adjudicated top-1 accuracy of 0.793 against the human labels. Sentiment was classified as negative, neutral, or positive using a fine-tuned Chinese RoBERTa model that achieved 0.8952 accuracy and a macro-F1 of 0.8858 on a review-disjoint test set. Images were classified into five categories using CLIP embeddings with logistic regression, reaching 0.800 accuracy and 0.949 top-2 accuracy on a held-out validation set of 215 images.

The visual corpus proved markedly more concentrated than the textual one. Building space accounted for 2,822 images, or 40.46 percent, and cultural symbols for 2,232 images, or 32.00 percent, while environment landscape made up 14.64 percent, people and activity context 8.62 percent, and service or access cues just 4.29 percent. The text, by contrast, distributed itself across all seven topics, with architectural heritage and spatial form the most frequent at 28.32 percent of sentences, followed by on-site process at 15.40 percent and environmental landscape at 13.53 percent. Notably, service- and access-related discourse was the only topic in which negative sentences outnumbered positive ones, at 52.01 percent versus 28.55 percent.

When the researchers examined how text and images corresponded within the same reviews, the picture became more nuanced. Among the 829 reviews containing both valid text and valid images, 787 contained at least one image classified as building space, cultural symbol, or environment landscape, and 375 of those, or 47.65 percent, also contained heritage-related discourse, meaning references to historical culture, scholarly atmosphere, or interpretation and knowledge. Concurrence was higher in multi-sentence reviews, at 54.66 percent, than in single-sentence reviews, at 40.51 percent, an odds ratio of 1.77. Yet even among multi-sentence reviews containing visible heritage imagery, 45.34 percent contained no heritage-related discourse at all, underscoring how partial the overlap between the two modalities really is.

Statistical testing revealed how fragile cross-modal associations can be. In bivariate analyses, four of the 35 topic-image combinations met the Benjamini-Hochberg false-discovery-rate criterion, including a positive association between architectural discourse and building-space images with an odds ratio of 2.57. But once the models accounted for academy, the number of sentences, and the number of images within each review, none of the 35 associations met the family-wide FDR criterion. The architectural-discourse and building-space estimate fell to an odds ratio of 1.55 with an adjusted q-value of 0.247, and a high-confidence sensitivity analysis using only the classifier’s most certain image labels produced similarly attenuated results. Cross-modal correspondence, the authors conclude, must be interpreted alongside site context and the extent of visitors’ contributions.

The evaluative analyses produced some of the study’s most telling findings. In a covariate-controlled model of 812 core matched reviews, heritage-related discourse was positively associated with complete-review textual valence, with a coefficient of 0.1212. But when the researchers removed the heritage-related sentences themselves and re-estimated the association using the remaining text in the same 617 reviews, the coefficient dropped to 0.0432 and was no longer statistically reliable. The paired coefficient difference of 0.098, with a 95 percent confidence interval of 0.049 to 0.147, indicates that heritage content and evaluative language were concentrated within the same textual segments: when visitors discussed historical culture, scholarly atmosphere, or interpretation, they tended to express approval or disappointment at the same time. Meanwhile, heritage-related discourse showed no statistically reliable association with platform star ratings at any of four cumulative-logit thresholds, and 72.98 percent of the 2,798 core reviews carried five stars, a concentration that limited the rating measure’s discriminatory range.

A separate content-coding measure captured explicit evaluations of environmental qualities, or EEQE, in which an eligible statement explicitly attributed a quality such as historical or academic atmosphere, tranquility, solemnity, restorative character, or artificiality to the academy or its immediate setting. Two authors independently coded 700 reviews, achieving pre-adjudication agreement of 0.9340 and a Cohen’s kappa of 0.8566. After adjudication, 447 of 699 determinate records contained an explicit environmental-quality evaluation, with 375 positive, 53 negative, 17 mixed-direction, and two neutral-descriptive cases. Crucially, these object-specific judgments did not always align with the overall evaluative tone of the review: among the 53 reviews with negative environmental-quality evaluations, 18 had positive complete-review valence, while among the 375 positive evaluations, 22 had zero valence and four had negative valence. Because the coding sample combined candidate-enriched records with a comparison sample, the authors caution that these counts characterize the coded sample rather than the full corpus, and the candidate-screening procedure’s estimated recall was only about 38 percent.

The authors frame the findings as evidence that visitor experience in educational heritage is a layered form of person-environment expression, in which photographs, heritage-related discourse, explicit environmental-quality appraisal, textual valence, and platform ratings each preserve a distinct informational role. They suggest that heritage managers could use these different indicators for different purposes, from identifying which material features enter public imagery to locating complaints about queues, transport, or pricing. The study’s scope carries clear limitations: Ctrip users and their content are self-selected, the analysis covers four Chinese academies on a single platform, text and image classification retain measurement error, and the post-visit design observes public expression at one point in time. The authors note that establishing how environmental exposure, understanding, and affect develop over time would require independent temporal or experimental evidence, and they call for future work combining multimodal records from additional platforms with environmental baselines and direct psychological measures.

Subject of Research: Multimodal analysis of visitor reviews and photographs from Chinese heritage academies

Article Title: Selective multimodal representation and visitor appraisal in educational heritage: evidence from Chinese academies

Article References: Niu, W., Zeng, Q., Liang, F., & Yu, Y. (2026). Selective multimodal representation and visitor appraisal in educational heritage: evidence from Chinese academies. Frontiers in Psychology, 17, Article 1926279. https://doi.org/10.3389/fpsyg.2026.1926279

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1926279

Keywords: educational heritage, Chinese academies, visitor reviews, multimodal user-generated content, environmental psychology, heritage tourism, text-image correspondence, sentiment analysis, star ratings, environmental quality evaluation, Ctrip, visitor appraisal