DH 2026

Daejeon, July 27–31

Fri, July 3111:00–12:30S066106
Short Paper

Knowledge-Augmented Generation (KAG) for Classical Literature: A Multilingual Chatbot Built on Structured Semantic Data of Sihwa ch’ongnim

Christina Han
Wifrid Laurier University, Canada · chan@wlu.ca
Lyndsey Twining
Independent Researcher · lyndseytwining@gmail.com
Jing Hu
Berline State Library · hu777jing@gmail.com
Yeong Won Chi
Korea University · chi123123@naver.com
Jonghoon Yoon
Arts Council Korea Artsarchive · hoonjong0821@gmail.com
TopicsEnglishartificial intelligence and machine learningdata modelingdatabase creationmanagementand analysislinked (open) datanatural language processingAsian studiesCultural studiesHumanities computingLiterary studiesTranslation studies
KeywordsKnowledge-Augmented Generation (KAG)Classical Sinitic LiteratureLinked Open Data (LOD)Multilingual Semantic Knowledge GraphsKorean Studies

This paper presents an ongoing project to build a multilingual, knowledge-grounded conversational agent for interpreting classical Sinitic literature using small but fully curated semantic data. The prototype is developed on top of a Semantic MediaWiki (SMW)–based Linked Open Data (LOD) archive constructed around the seventeenth-century Korean anthology Sihwa ch’ongnim 詩話叢林 (Compendium of Poetry Talks) [Han et al. 2022]. Rather than serving as a translation engine or text-retrieval tool, the system explores how knowledge-augmented generation (KAG)—an emerging paradigm that integrates LLM outputs with structured graph reasoning—can support translingual interpretation in ways conventional retrieval-augmented generation (RAG) cannot [Gao et al. 2023; Han et al. 2024; Jeong 2024; Li, Z. et al. 2025; Liang et al. 2025].

We argue that classical East Asian corpora, with their multilingual ecosystems (classical Sinitic, modern Korean, modern Chinese, and English scholarship), provide a crucial testbed for rethinking translingual reasoning in AI. General-purpose LLMs struggle to navigate the philological and cultural depth of classical Sinitic texts, demonstrating the need for systems that combine multilingual capacity with structured knowledge representations. We aim to demonstrate that curated, multilingual LOD can meaningfully enhance LLM interpretation of premodern literature.

From RAG to KAG

Recent AI work on classical literature has concentrated on text mining, RAG pipelines, and domain-specific LLMs. Korean scholarship on literary chatbots notes that while RAG offers strong retrieval and accessible querying, it struggles with contextual reasoning, maintaining knowledge accuracy, and generating meaningful interpretive insights [Ha & Park 2025; Jeong 2024]. These weaknesses become more pronounced with classical Sinitic texts, such as those in the Sihwa ch’ongnim anthology, where compact syntax, dense allusions, and culturally layered references challenge generalist LLMs trained on modern data [Li, Z. et al. 2025; Liang et al. 2025].

Domain-specialized models such as TongGu and SikuGPT improve comprehension through targeted pretraining [Cao et al. 2024; Liu et al. 2023], yet they remain mostly monolingual and rely on unstructured text retrieval. Without structured knowledge graphs, they cannot perform explicit reasoning over the relational features essential to deep literary interpretation—such as links among people, places, genres, and historical events [Ma et al. 2020; Liu et al. 2025].

Examples from recent scholarship illustrate the limitations of retrieval-only methods: RAG pipelines for classical literature often retrieve relevant passages but fail to maintain coherence or capture cultural nuance [Jeong 2024; Ha & Park 2025]. Work on Yanxinglu (Records of Missions to China) similarly shows difficulty representing spatial and narrative structure using unstructured retrieval alone [Chen & Shi 2024]. These shortcomings stem from RAG’s dependence on similarity metrics rather than structured reasoning.

Our project adopts a KAG approach to address these limitations. KAG integrates structured knowledge graphs derived from LOD to actively inform retrieval and generation, moving beyond similarity-based matching. Specifically, we implement this using GraphRAG-style reasoning. The system translates conversational queries into graph traversals and structured reasoning, allowing scholars to engage with relational data through natural language while avoiding the technical barriers of SPARQL interfaces [Barbera 2013; Middle 2022]. KAG also enables entity alignment across languages and scripts within the knowledge graph—an essential capability for multilingual data like the Sihwa ch’ongnim LOD [Spence & Brandao 2021; Horvath 2021].

Developing KAG with Structured Data: The Sihwa ch’ongnim LOD

The Sihwa ch’ongnim provides an ideal small-data testbed for KAG because its interpretive depth depends on relational context—how people, places, occasions, and poetic concepts interact within each narrative. Each sihwa entry combines anecdotal prose with embedded poems and critical remarks, documenting poetic composition, exchange, and evaluation in premodern Korea. These relational dynamics are difficult for unstructured text retrieval to capture but align naturally with a structured LOD approach.

Our SMW platform encodes the entire anthology into a detailed semantic schema, including 932 sihwa anecdotes, 1,222 historical figures, and 547 geo-referenced locations. All entities are multilingual (Sinitic, Korean, English) and linked to external LOD sources such as Wikidata. This structured environment enables the system to connect classical texts with their translations, biographical data, and broader historical context.

Because the archive is expressed as a network of RDF triples, the KAG system can reason directly over the relationships among texts, people, places, terms, and events. We expect this will enable the chatbot to follow narrative context, recognize social relationships, interpret poetic terminology, and align entities across multiple languages with greater precision. Everyday questions are automatically converted into graph queries, producing responses grounded in curated semantic data.

Case Study: KAG in the Sihwa ch’ongnim Context

Currently, we have linked our KAG-based LLM prototype to a Neo4j graph database with our data and are in the process of testing and refining the prototype. By the time of the short paper presentation at the conference, we hope to demonstrate 1) successful semantic data query through natural-language prompts in multiple languages, 2) a live comparison of KAG versus RAG-only query results, 3) functionality across the various scripts in our database (Sinitic, Korean, and English), and 4) citation-grounded responses that point back to specific URIs within the semantic database.

Below are screenshots that demonstrate an application use case in which a complex natural-language question is converted by a KAG chatbot into a semantic query (Cypher query), queried in Neo4j, and then—after extracting the data—a response is provided to the user.

Above, the chatbot responds to the question “What poem did Kim Sinyun compose in the kyŏngin year of King Ŭijong’s reign?” in Korean with the correct poem in its original Sinitic and its Korean and English translations.

Conclusion

This project positions KAG as a powerful framework for interpreting classical Sinitic literature using the Sihwa ch’ongnim LOD. By leveraging our curated, multilingual LOD, the system moves beyond RAG’s similarity-based retrieval to achieve richer, context-aware interpretation and cross-lingual entity alignment, essential for complex premodern texts. Our work demonstrates that interpretable AI for classical literature depends on meticulously structured knowledge representation rather than massive unstructured corpora. This approach is scalable to other cultural heritage domains, offering a new path for integrating advanced AI with curated humanities data.

References
  1. Barbera, M. (2013): “Linked (open) data at web scale: research, social and engineering challenges in the digital humanities”, in: JLIS.It: Italian Journal of Library and Information Science 4, 1: 91–104. DOI: 10.4403/jlis.it-6333.
  2. Cai, Z. / Kurzynski, M. (2024): “Brevity and Breadth: A Linguistic, Aesthetic, and DH-Assisted Study of the Book of Poetry and 'Nineteen Old Poems'”, in: Journal of Chinese Literature and Culture 11, 2: 235–264. DOI: 10.1215/23290048-11410394.
  3. Cao, J. / Shi, Y. / Peng, D. / Liu, Y. / Jin, L. (2024): “C 3 Bench: A Comprehensive Classical Chinese Understanding Benchmark for Large Language Models”, in: Computation and Language (May 20). DOI: 10.48550/arxiv.2405.17732.
  4. Cao, J. / Peng, D. / Zhang, P. / Shi, Y. / Liu, Y. / Ding, K. / Jin, L. (2024): “TongGu: Mastering Classical Chinese Understanding with Knowledge-Grounded Large Language Models”, in: Computation and Language (September 30). DOI: 10.48550/arxiv.2407.03937.
  5. Chen, A. / Lou, L. / Chen, K. / Bai, X. / Xiang, Y. / Yang, M. / Zhao, T. / Zhang, M. (2024): “Large language models for classical Chinese poetry translation: Benchmarking, evaluating, and improving”, in: Computation and Language (December 30). DOI: 10.48550/arXiv.2408.09945.
  6. Chen, A. / Lou, L. / Chen, K. / Bai, X. / Xiang, Y. / Yang, M. / Zhao, T. / Zhang, M. (2025): “Benchmarking LLMs for translating classical Chinese poetry: Evaluating adequacy, fluency, and elegance”, in: Association for Computational Linguistics (ed.): Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: 33008–33025. DOI: 10.18653/v1/2025.emnlp-main.1678.
  7. Chen, J. / Shi, H. (2024): ‘Construction of LLM–RAG pipeline based on spatial narrative characteristics of Yanxinglu”, in: Segae Hanja Yŏn'gu 7, 2: 43–70.
  8. Chi, Y. W. / Choi, J. K. (2024): “A conceptual data modeling attempt for building a Korean classical poetry database”, in: Minjok Munhwasa Yŏn'gu 85: 43–82.
  9. De Weerdt, H. / Ho, H. I. / Simon, R. / Lee, S. / Molenaar, S. / Xi, W. / Zhuang, D. / Stojević, I. / Tu, H.-C. / Zaneri, T. / Lin, N.-Y. / Meister, M. (2025): “Contextual semantic text and image annotation in the MARKUS environment”, in: Digital Humanities Quarterly 19, 4 <https://dhq.digitalhumanities.org/vol/19/4/000808/000808.html> [03.05.2026].
  10. Digital Victorian Periodical Poetry (n.d.): Digital Victorian Periodical Poetry <https://dvpp.uvic.ca/index.html> [03.05.2026].
  11. Doh, W. Y. (2025): “The necessity and direction of compiling a dictionary of terms in Korean classical Chinese texts”, in: Minjok Munhwa 69: 117–146.
  12. Encyclopedia of Romantic Nationalism (n.d.): Encyclopedia of Romantic Nationalism <https://ernie.uva.nl/viewer.p/21/56> [03.05.2026].
  13. Gao, Y. / Xiong, Y. / Gao, X. / Jia, K. / Pan, J. / Bi, Y. / Dai, Y. / Sun, J. / Wang, M. / Wang, H. (2023): “Retrieval-Augmented Generation for Large Language Models: A Survey”, in: Computation and Language (March 27). DOI: 10.48550/arxiv.2312.10997.
  14. Ha, D. J. / Park, M. (2025): “Text mining analysis of jingpai and haipai literary works: Possibilities and limitations of RAG-based chatbot analysis”, in: Simininmunhak 49: 203–231. DOI: 10.22842/kgucfh.2025.49.203.
  15. Han, C. / Chi, Y. W. / Hu, J. / Ryu, I. T. (2022): “A foundational design for creating a sihwa semantic data archive”, in: Hanmunhak Nonjip 63: 105–146. DOI: 10.17260/jklc.2022.63.105.
  16. Han, H. / Wang, Y. / Shomer, H. / Guo, K. / Ding, J. / Lei, Y. / Halappanavar, M. / Rossi, R. A. / Mukherjee, S. / Tang, X. / He, Q. / Hua, Z. / Long, B. / Zhao, T. / Shah, N. / Javari, A. / Xia, Y. / Tang, J. (2024): “Retrieval-Augmented Generation with Graphs (GraphRAG)”, in: Computation and Language (January 8). DOI: 10.48550/arxiv.2501.00309.
  17. Horvath, A. (2021): “Enhancing language inclusivity in Digital Humanities: Towards sensitivity and multilingualism”, in: Modern Languages Open 1. DOI: 10.3828/mlo.v0i0.382.
  18. Hou, Y. / Frank, A. (2015): “Analyzing sentiment in classical Chinese poetry”, in: Association for Computational Linguistics (ed.): Proceedings of the 9th SIGHUM Workshop on Language Technology for Cultural Heritage, Social Sciences, and Humanities (LaTeCH), July 2015: 15–24. DOI: 10.18653/v1/W15-3703.
  19. Jeong, C. (2024): “A graph-agent–based approach to enhancing knowledge-based QA with advanced RAG”, in: Chisik Kyŏngyŏng Yŏn'gu 25, 3: 99–119.
  20. Kang, C. (2025): “ChatGPT utilization in the study and education of Geumo Sinhwa: With critical reflections”, in: Ŏmun Ronch'ong 103: 63–90.
  21. Lee, B. C. (2024): “Application and limitations of text mining techniques in classical Sinitic poetry”, in: Ŏmun Yŏn'gu 121: 219–249.
  22. Lee, G. H. / Byun, E. M. / Ryu, I. T. (2024): “Semantic data processing of civil service examination materials in the Chosŏn Dynasty”, in: Han'gukhak Munhwa Yŏn'gu 92: 65–104.
  23. Lee, S. E. (2024): “The genealogy of stories: Reading success narratives in yadam through digital humanities methodology”, in: Journal of Korean Culture 66: 143–180.
  24. Li, S. / Hu, R. / Wang, L. (2025): “Efficiently Building a Domain-Specific Large Language Model from Scratch: A Case Study of a Classical Chinese Large Language Model”, in: Computation and Language (June 19). DOI: 10.48550/arxiv.2505.11810.
  25. Li, Z. / Wang, Z. / Wang, W. / Hung, K. / Xie, H. / Wang, F. L. (2025): "Retrieval-augmented generation for educational application: A systematic survey", in: Computers and Education. Artificial Intelligence 8, Article 100417. DOI: 10.1016/j.caeai.2025.100417.
  26. Liang, L. / Bo, Z. / Gui, Z. / Zhu, Z. / Zhong, L. / Zhao, P. / Sun, M. / Zhang, Z. / Zhou, J. / Chen, W. / Zhang, W. / Chen, H. (2025): “KAG: Boosting LLMs in professional domains via Knowledge Augmented Generation”, in: Companion Proceedings of the ACM on Web Conference 2025: 334–343. DOI: 10.1145/3701716.3715240.
  27. Lin, S. (2020): “The Study of Premodern Chinese Literature in the Digital Era: New Methods of Quantitative Statistics, Databases, and Visualization Analyses”, in: Library Trends 69, 1: 269–288. DOI: 10.1353/lib.2020.0032.
  28. Linked Art (n.d.): “LOUD: Linked Open Usable Data” <https://linked.art/loud/> [03.05.2026].
  29. Liu, C. / Wang, D. / Zhao, Z. / Hu, D. / Wu, M. / Lin, L. / Liu, J. / Zhang, H. / Shen, S. / Li, B. / Zhao, L. (2024): “SikuGPT: A Generative Pre-trained Model for Intelligent Information Processing of Ancient Texts from the Perspective of Digital Humanities”, in: Journal on Computing and Cultural Heritage 17, 4: Article 53. DOI: 10.1145/3676969.
  30. Liu, C. L. / Mazanec, T. J. / Tharsen, J. R. (2018): “Exploring Chinese poetry with digital assistance: Examples from linguistic, literary, and historical viewpoints”, in: Journal of Chinese Literature and Culture 5, 2: 276–321. DOI: 10.1215/23290048-7257002.
  31. Liu, Z. / Wan, G. / Zuo, X. / Liu, Y. (2025): “Sentiment analysis of Chinese ancient poetry based on multidimensional knowledge attention”, in: Digital Scholarship in the Humanities 40, 1: 214–226. DOI: 10.1093/llc/fqae069.
  32. Living Poets (n.d.): Living Poets <https://livingpoets.dur.ac.uk/w/index.php/Welcome_to_Living_Poets> [03.05.2026].
  33. Lv, W. / Cao, Q. / Liu, X. (2025): “A multi agent classical Chinese translation method based on large language models”, in: Scientific Reports 15, 1: Article 40160. DOI: 10.1038/s41598-025-23904-0.
  34. Ma, B. / Yao, Y. / Haensch, A.-C. (2025): “Capabilities and evaluation biases of large language models in classical Chinese poetry generation: A case study on Tang poetry”, in Computation and Language (April 20). DOI: 10.48550/arXiv.2510.15313.
  35. Ma, Z. / He, J. / Liu, S. (2020): “Representation of the spatio-temporal narrative of The Tale of Li Wa”, in: PloS One 15, 4: e0231529. DOI: 10.1371/journal.pone.0231529.
  36. Mapping the Republic of Letters (n.d.): Mapping the Republic of Letters <http://republicofletters.stanford.edu/> [03.05.2026].
  37. Middle, S. (2021): Investigating Linked Data Usability for Ancient World Research. ProQuest Dissertations & Theses.
  38. Si, Z. (2026): “Mapping the spatial–temporal evolution of imagery in Tang poetry: A computer vision and GIS-based approach”, in: Future Digital Technologies and Artificial Intelligence 2, 1: 1–6.
  39. Spence, P. J. / Brandao, R. (2021): “Towards language sensitivity and diversity in the Digital Humanities”, in: Digital Studies 11, 1. DOI: 10.16995/dscn.8098.
  40. Yao, X. / Wang, M. / Chen, B. / Zhao, X. (2025): “WenyanGPT: A Large Language Model for Classical Chinese Tasks”. DOI: 10.48550/arxiv.2504.20609.
  41. Yoo, J. J. (2024): “A preliminary study on the construction of an ontology for the study of classic Sinitic poetry in late Chosŏn Korea: Focusing on the function of inference”, in: Hangukhak Nonjip 96: 5–41.
  42. Yu, H. / Gao, C. / Li, X. / Zhang, L. (2024): “Ancient Chinese Poetry Collation Based on BERT”, in: Procedia Computer Science 242: 1171–1178. DOI: 10.1016/j.procs.2024.08.179.
  43. Zhang, Q. / Chen, S. / Bei, Y. / Yuan, Z. / Zhou, H. / Hong, Z. / Chen, H. / Xiao, Y. / Zhou, C. / Dong, J. / Chang, Y. / Huang, X. (2025): “A Survey of Graph Retrieval-Augmented Generation for Customized Large Language Models”, in: Computation and Language (September 29). DOI: 10.48550/arxiv.2501.13958.
  44. Zhang, W. / Lai, Z. / Tang, S. (2025): “Spatiotemporal distribution characteristics of Nanjing place names—Based on data mining of Tang-Song poetry and online travelogues”, in: PloS One 20, 2: e0319244. DOI: 10.1371/journal.pone.0319244.
  45. Zhang, W. / Wang, H. / Song, M. / Deng, S. (2023): “A method of constructing a fine-grained sentiment lexicon for the humanities computing of classical Chinese poetry”, in: Neural Computing & Applications 35, 3: 2325–2346. DOI: 10.1007/s00521-022-07690-8.
  46. Zheng, M. / Moeller, S. (2025): “Challenges in processing Chinese texts across genres and eras”, in: Association for Computational Linguistics (ed.): Proceedings of the 9th Widening NLP Workshop, November 2025: 230–234.
  47. Zou, D. / Chen, Y. / Li, M. / Miao, S. / Liu, C. / Han, B. / Cheng, J. / Li, P. (2025): “Weak-to-Strong GraphRAG: Aligning Weak Retrievers with Large Language Models for Graph-based Retrieval Augmented Generation”, in: Computation and Language (June 26). DOI: 10.48550/arxiv.2506.22518.
PreviousArtificial Memory as Compass: Navigating Collective Knowledge in the Age of AINextWhat archives remember, let Triples render! Critical modeling of historical knowledge as linked (open) data