Associations among park features, physical activities, and sensory perceptions from online reviews: A domain-specific named entity recognition model

Research on human–environment interactions remains fragmented, with limited exploration of how diverse park features jointly relate to multiple physical activity (PA) levels and sensory perceptions. Large-scale textual data offer new opportunities to capture public experiences of parks. However, prevailing natural language processing approaches, such as lexicon-based and prompt-based methods, often lack contextual sensitivity and accuracy. To address these limitations, we developed a landscape character named entity recognition (LCNER) model fine-tuned on a manually curated dataset to simultaneously extract park features, sensory perceptions, and three PA levels from online reviews. Among six pre-trained language models evaluated, the DeBERTa-large–based LCNER achieved the highest mean F1 score (0.896 ± 0.001) and outperformed domain-lexicon baselines, with the largest improvements observed for the best-performing entity categories: ＋0.558 in precision, ＋0.231 in recall, and ＋0.395 in F1 score. Quasi-binomial analyses revealed that facility-related features provided better model fit for moderate-to-vigorous PA (MVPA) and sound perception than for other activities and sensory types. Several features exhibited opposite associations with MVPA and sedentary behavior. Moreover, certain activity-oriented facilities were negatively associated with sensory perceptions, suggesting a potential trade-off between active engagement and sensory awareness. Overall, LCNER demonstrates the potential to unify the extraction of park features, activity levels, and sensory perceptions from online texts, advancing understanding of how park environments shape human experiences and behaviors. Code and prompts are available at https://github.com/Wenpeimuzi/Landscape-NLP.