General-purpose large models for natural language processing of the Han Fu depicting architecture: a performance evaluation

Ancient texts, including the Han Fu depicting architecture, are vital for studying cultural and architectural history. Translating them into English facilitates cultural exchange. Large language models show promise for revitalizing Han Fu, but vertical domain large models (VSLMs) in this field have limited functions, and the capabilities of general-purpose large models (GPLMs) remain underexplored. This study evaluates GPLMs’ performance in translating architectural descriptions in Han Fu, using Yang Xiong’s Ganquan Fu as an example. The research employed a three-stage framework: (1) When 13 GPLMs translated 163 architectural terms, DeepSeek V3, Hunyuan-T1, and ERNIE Bot 4.5 Turbo ranked top via expert scoring. (2) Sentence segmentation and structural analysis showed that GPLMs achieved high punctuation accuracy (Dice Coefficient: 0.89 – 0.92), comparable to that of VSLM AI Taiyan, but struggled with intra-sentential relation extraction (Dice Coefficient: 0.21 – 0.30). (3) In the full-text translation from vernacular Chinese to English, which was evaluated via bilingual corpus alignment, ERNIE Bot 4.5 Turbo delivered the most complete versions, though architectural terminology required manual refinement. The findings highlight GPLMs’ potential in basic tasks but reveal limitations in semantic consistency and specialized knowledge transfer, emphasizing the need for hybrid human-AI workflows in revitalizing classical architectural literature.