Naoya Iwata and Ikko Tanaka Presented at the UTokyo Digital Humanities Workshop

On February 23, 2026, Naoya Iwata (Nagoya University / National Institute of Informatics) and Ikko Tanaka (J. F. Oberlin University), both core members of the Humanitext project, delivered presentations at the Digital Humanities Workshop titled “Connecting Resources and Knowledge with AI.” The event was hosted by the Center for the Development of Next-Generation Humanities at the University of Tokyo and took place both online and at the UTokyo Yaesu Academic Commons.

The workshop aimed to explore the potential of AI in Digital Humanities (DH), focusing on interdisciplinary discussions regarding the integration of research materials and knowledge bases using Retrieval-Augmented Generation (RAG), as well as the structuring of historical information via Knowledge Graphs.

Toward Humanities-Specific RAG Evaluation Metrics (Naoya Iwata)

In the first discussion session, “The Potential of RAG in Digital Humanities,” Naoya Iwata presented on “How to Evaluate Search Accuracy in Humanities RAG: Insights from a Pilot Evaluation on Western Classical Texts.”

While demonstrating the improved search accuracy achieved through Context-Oriented Translation (COT) and Hybrid Search in Humanitext Antiqua, Iwata pointed out a critical issue: standard RAG evaluation metrics (such as RAGAS) are not necessarily suitable for the humanities. In humanistic research, there is rarely a single “correct answer,” and texts possess complex hierarchical structures consisting of original sources, commentaries, and fragments.

To address this, the research team conducted a pilot evaluation with experts and proposed unique evaluation axes tailored for humanities RAG: Source Stratification, Source Type Sensitivity, and Result Diversity. Iwata emphasized the necessity of moving away from flat data structures and instead adopting a three-tier architecture comprising a TEI foundational layer, a text reference layer, and a knowledge graph reasoning layer.

Building Networks of Interpretation Connecting “Originals” and “References” (Ikko Tanaka)

During the second discussion session, “Connecting Primary Sources with Related Information and Context,” Ikko Tanaka gave a presentation titled “Thinking About ‘Original Texts’ and ‘References’: Building a System to ‘Interpret’ Western Classics.”

Tanaka reported on the profound difficulties of establishing “references” encountered while developing the Humanitext Reader, a text-reading support environment for Western Classics. He illustrated that the task of creating links is not merely technical but deeply conceptual, entangled with issues such as the varying granularity of references, fluctuations in the concept of an “original text” due to different editions, unexplicit mentions in Running Commentaries, and references between secondary literatures.

The act of stating “Text A refers to Text B” relies on a tradition of interpretation and critique spanning thousands of years. Tanaka advocated for the creation of an “explorable network” where multiple interpretations can coexist and the structure dynamically adapts to the researcher’s inquiries, rather than a static collection of unidirectional links. He concluded his presentation by sharing the vision for “Humanitext GEO,” an upcoming spatial interface designed to connect time, space, and texts.

Building on the discussions from this workshop, the Humanitext project will continue to advance the development of next-generation research infrastructures that seamlessly integrate cutting-edge AI technologies with rigorous humanistic scholarship.