Visual business cards provide a digital medium for presenting information on intangible cultural heritage (ICH) inheritors. Accurately recognizing these cards allows the extraction of key details that, when combined with large language models, support semantic understanding and contextual reasoning to generate enriched cultural content. This study introduces a Graph-Retrieval framework that integrates graph-based methods with retrieval-augmented generation for visual information recognition and ICH-related question answering. A dataset of Chinese ICH inheritors’ visual cards was built, covering 10 information types. To enhance extraction robustness, graph feature enhancement is applied through semantic recognition, random node masking, and edge deletion, while positional attention captures spatial relationships. A complementary Loop-RAG strategy dynamically integrates external knowledge with inner-outer loop retrieval. Experiments show the Graph-Retrieval framework surpasses benchmarks on multiple datasets, achieving a macro-average F1 of 0.928, with ablation studies validating feature enhancements. Loop-RAG also excels in report generation, question answering and few-shot conditions.
Wang et al. (Fri,) studied this question.