AI-Mediated Knowledge Graphs for Distant Reading of Urban Perception Abstract This study advances urban spatial perception research by moving beyond conventional physical indicators to foreground public subjective experience, addressing the spatial bias and limited interpretability that characterize approaches reliant on user-generated content or static visualization. We propose a framework that integrates multimodal large language models with knowledge graphs within the digital humanities paradigm of distant reading, enabling large-scale, structural interpretation of urban perception narratives. Using Beijing, a city rich in heritage and rapid modernization, as a case study, we generated structured perceptual texts from Baidu street-view imagery with an LLM and operationalized thirty perceptual indicators. After standardization, hierarchical clustering, DBSCAN, and K-means were used to segment the corpus into high, medium, and low perception categories. Texts underwent tokenization, domain dictionary mapping, and synonym normalization, followed by construction of co-occurrence matrices with a sliding window. Semantic networks were visualized in VOSviewer and Pajek and analyzed with community detection and centrality measures. The resulting knowledge graph reveals a multilayered semantic architecture in which core clusters of physical and spatial descriptors such as street view, architecture, and vegetation provide a descriptive backbone, while peripheral clusters extend into environmental and functional dimensions such as ecology and vitality and into aesthetic and emotional registers such as artistic and diverse; low-perception texts foreground discourses of lack and ordinariness, and marginal modules expose interpretive limits of LLMs under ambiguous imagery or contextual uncertainty. This framework advances digital humanities by demonstrating how distant reading of urban perception texts—operationalized through AI-mediated generation, clustering, and semantic graphing—can reveal macro-level patterns of how urban heritage spaces are cognitively framed. Conceptually, the framework treats LLM-generated perceptual texts as cultural artifacts that crystallize collective imaginaries of heritage cities, and practically it offers heritage-sensitive insights for urban planning by rendering perceptual contrasts and methodological boundaries explicit.