Unfolding the multisensory landscape experiences from online reviews: A large language model based approach with implications for human-centric land use policy

Understanding multisensory landscape experiences has recently attracted considerable interest in the realm of land use policy, given the crucial role of five external senses in people’s spatial perception, behavior and memory, thereby offering new chances for incorporating human inputs in spatial planning. However, two methodological challenges remain unsettled for examining human multisensory landscape experiences. For one thing, sensory experiences are context-dependent and neglecting the context would result in inaccuracies and incompleteness in measurement. For another, multisensory landscape experiences co-occur in a complex manner and it is quite difficult to characterize their synergy variations across places. This paper proposes a large language model based approach to unfolding multisensory landscape experiences from online reviews. We first construct a context-sensitive ChatGLM3–6B model, fine-tuned for multisensory landscape experiences triplet extraction under the instruction learning paradigm, utilizing a custom dataset designed to efficiently recognize sensory descriptions, idiomatic expressions, and emotional symbolism. Based on the measurements of multisensory experiences, a new probabilistic Bayesian model, the 'Discrete Geographical Topic Model', is developed to facilitate the identification of multisensory synergy related to specific geographic sites. We validate our approach using the case of Jiangnan Classical Gardens, a group of quintessential Chinese cultural heritage sites. Results demonstrate that the approach can efficiently extract multisensory landscape experiences using limited training data. Based on the findings, we finally discuss some essential implications for human-centric planning. Our study foregrounds the enormous promise of large language models in advancing the human-centric practices in land use policy.