College of Arts and Sciences Ikko Tanaka Humanitext, with Professor Participation, Selected for the International Research Program HAVI

The research project "Humanitext," involving Professor Ikko Tanaka J. F. Oberlin University College of Arts and Sciences, has been selected for the international research program "Humanities and AI Virtual Institute (HAVI)" by the U.S. nonprofit Schmidt Sciences. Out of 501 applications submitted from 42 countries worldwide, this is one of 19 selected.
This project is led by Associate Professor Naoya Iwata of the Center for Promotion of Digital Humanities and Social Sciences at Nagoya University, and is jointly advanced by researchers from our university, the University of Tokyo, and the National Institute of Informatics. AI will propose relationships such as citation, criticism, and rephrasing among classical literature, and experts will verify them, thereby constructing a knowledge graph that traces the connections between texts along with evidence.

The official name is "Humanitext: A Human-AI Co-Creation Platform for Verified Classical Scholarship." The total grant amount is US$816,667 (approximately 128 million yen) over three years, from October 2026 to September 2029. The grant will be disbursed in two 18-month periods, with the second period's disbursement scheduled to follow a progress evaluation.

Professor Tanaka will leverage his expertise in Greek philosophy to work on the interpretation of classical texts and the verification of relationships between them. He will be responsible for quality control of annotations, including creating guidelines for annotation work and evaluating the degree of agreement among commentators, while also designing a knowledge graph database and conducting final verification as an expert.

This project will develop a system based on Humanitext Antiqua where AI and experts collaborate to build relationships between texts. First, researchers will manually construct 500 relationships from genres such as philosophical commentaries, historical narratives, and literary allusions, using these as a basis for verifying and evaluating the data design. Then, the AI will propose new relationship candidates, and a team of experts and graduate students will decide to accept, revise, or reject each one.
The verified relationships are stored as a knowledge graph that records not only their type (such as "quote," "allegorization," "paraphrase," and "criticism"), but also the direction of the relationship, its likelihood, and the history of the verification. Wikibase and the graph database Neo4j are used to manage the data.

Furthermore, we will use "GraphRAG," a method that utilizes verified relationships as the basis for AI responses, to evaluate how much the reliability of responses and the accuracy of citations improve compared to conventional search augmentation generation (RAG). We also plan to collaborate with external classical scholars in the evaluation. We aim to build more than 10,000 verified relationships over three years and expand to Latin-language texts in the third year.

The knowledge graph, annotation software, and evaluation benchmark data that we have constructed will be released under an open license. The functions based on the knowledge graph will also be made available to the general public through Humanitext. This is expected to enable everyone to examine, based on data, questions such as which arguments have been debated among commentators and how ideas have changed across time and tradition.
Furthermore, a system in which experts verify AI suggestions and use the results to improve the AI is expected to have applications in specialized fields where accuracy is crucial, such as medical literature and law.

HAVI is an international research program run by Schmidt Sciences that promotes collaboration between humanities and AI researchers to advance both fields. In 2026, it will provide a total of US$10 million in grants to 19 research teams, bringing the total grants for the past two years to over US$21 million. It supports interdisciplinary research exploring human history, culture, and expression, and places great emphasis on making developed code, models, and benchmark data widely available.

Schmidt Sciences is a U.S. non-profit organization founded in 2024 by former Google CEOs Eric and Wendy Schmidt. It works to accelerate scientific knowledge and breakthroughs in fields such as AI and advanced computing, astrophysics, life sciences, climate, and space, by utilizing cutting-edge tools.