Study Proposes Traditional Culture Generation Method

A study by Zi-Qi Liu in Discover Artificial Intelligence (2026) proposes a generative AI-driven method for improving large language model content about Chinese traditional culture. The method aims to address outdated knowledge, cultural-context mismatch, weak links to evidence, and unsupported additions. It keeps Qwen2.5-7B-Instruct frozen and connects it to an external knowledge base that can be updated, without changing the model’s internal attention layers.

The system labels knowledge with four fields: cultural category, historical stage, regional origin, and application scenario. It ranks candidate evidence using vector similarity, graph relations, tag consistency, source credibility, and penalties for conflicting information, then assembles the selected material into a constrained prompt. Knowledge updates combine source verification, conflict detection, and expert review. After generation, the system checks factual consistency, cultural-context adaptability, and evidence coverage.

Experiments included classical texts, cultural heritage objects, intangible cultural heritage crafts, traditional Chinese opera, and folk culture, using 1,080 factual question-answering samples and 360 contextual generation tasks. The proposed system reported 88.61% factual accuracy, 95.28% answer relevance, 94.17% evidence faithfulness, and 92.50% cultural-context adaptability. For questions dependent on updated knowledge, accuracy rose from 58.67% to 91.33% after updating. The authors describe the pipeline as traceable and context-sensitive. They interpret differences between systems descriptively because complete paired baseline predictions were not retained.

A study by Zi-Qi Liu in Discover Artificial Intelligence (2026) proposes a generative AI-driven method for improving large language model content about Chinese traditional culture. The method aims to address outdated knowledge, cultural-context mismatch, weak links to evidence, and unsupported additions. It keeps Qwen2.5-7B-Instruct frozen and connects it to an external knowledge base that can be updated, without changing the model’s internal attention layers.

The system labels knowledge with four fields: cultural category, historical stage, regional origin, and application scenario. It ranks candidate evidence using vector similarity, graph relations, tag consistency, source credibility, and penalties for conflicting information, then assembles the selected material into a constrained prompt. Knowledge updates combine source verification, conflict detection, and expert review. After generation, the system checks factual consistency, cultural-context adaptability, and evidence coverage.

Experiments included classical texts, cultural heritage objects, intangible cultural heritage crafts, traditional Chinese opera, and folk culture, using 1,080 factual question-answering samples and 360 contextual generation tasks. The proposed system reported 88.61% factual accuracy, 95.28% answer relevance, 94.17% evidence faithfulness, and 92.50% cultural-context adaptability. For questions dependent on updated knowledge, accuracy rose from 58.67% to 91.33% after updating. The authors describe the pipeline as traceable and context-sensitive. They interpret differences between systems descriptively because complete paired baseline predictions were not retained.