TCFP-Sight: Ontology-Gated DAIE Reasoning for Auditable Appreciation of Traditional Chinese Figure Painting

TCFP-Sight traces the DAIE reasoning path over Fisherman Resting on Autumn River (Ma Yuan, Southern Song): from L1/L2 visual evidence → Focus Set (D) → formal-language analysis (A) → theme–mood–value interpretation (I) → cultural-historical evaluation (E), producing a fully auditable appreciation report.

Abstract: Enabling artifi cial intelligence to transcend surface-level visual perception and achieve deep cultural-semantic understanding of Chinese fi gure painting is a core open challenge in Cultural Heritage Computing. Although multimodal large language models (MLLMs) can generate fl uent art commentary, they systematically fail at cross-level cognitive transitions—producing conclusions that are semantically coherent yet lack grounding in low-level visual evidence—a failure mode we term cultural hallucination. Existing retrieval-augmented generation (RAG) approaches, moreover, lack stage-structured reasoning constraints and domain-specific ontological support. We therefore present TCFP-Sight (Traditional Chinese Figure Painting Sight), an ontology-gated multi-stage reasoning framework for auditable appreciation of Chinese figure painting. The system is grounded in a structured prior knowledge space integrating Panofskyan iconology with Chinese painting theory into a computable semantic schema, instantiated across a knowledge graph of 1,323 annotated fi gure paintings spanning the Tang through Qing dynasties. An Ontology Gating mechanism prevents premature activation of high-level cultural meaning before low-level evidential foundations are secured, mandating that all output claims carry explicit evidence links and confi dence scores. Evaluated on a newly constructed benchmark of 100 classical paintings, TCFP-Sight achieves consistent gains over generic MLLMs and RAG baselines across appreciation quality and hallucination suppression dimensions, while improving evidence traceability. Ablation studies further confi rm that Ontology Gating is the most critical anti-hallucination component, and that the staged reasoning pipeline is an indispensable complement to domain knowledge-driven architectures.

Recommended citation: Xinguo Xu, Chang Liu, Xiaoying Xu, Xingjian Chai, Yuhong Ying, Huiqi Hou, Guodao Sun, Jiazhou Chen, Pengyi Hao, Fuli Wu. “TCFP-Sight: Ontology-Gated DAIE Reasoning for Auditable Appreciation of Traditional Chinese Figure Painting.” ACM Multimedia. 2026.