Longitudinal digital heritage perception at Ciqikou: sentiment analysis and time-sliced BERTopic of Dianping reviews

Mountainous port-town morphology provides the architectural context for this study. Focusing on Ciqikou, a riverine heritage settlement in Southwest China, we analyze eight years of Dianping reviews as a longitudinal record of platform-mediated heritage perception rather than physical change or all visitor experience. We examined 22,339 reviews using human-validated MacBERT sentiment classification and fixed BERTopic assignments aggregated by quarter. Two independent domain-background annotators labelled 8,572 reviews, with professor adjudication (raw agreement 85.72%; Cohen’s κ=0.7783, 95% CI 0.7672–0.7893). The independent test set achieved 0.967347 accuracy and 0.958317 macro-F1. BERTopic identified 136 non-outlier micro-topics and 2,446 outliers; G1–G19 are hierarchy-informed, domain-guided meso-level interpretations. A five-seed comparison of LDA, NMF, BERTopic and CombinedTM found no metric-independent winner, while BERTopic micro-topics were parameter-sensitive (ARI 0.268985–0.353060). Changepoint analysis of 31 complete quarters identified transitions at 2019Q3, 2020Q3, 2022Q1 and 2023Q2, with boundary-quarter sensitivity. The findings support heritage interpretation and visitor-experience monitoring while requiring caution about topic stability, breakpoint certainty, physical carrying capacity and causal inference.