仏教学とデジタル・ヒューマニティーズの部

13:00-14:30 デジタル中国仏教研究の最先端 

Professor Christoph Anderl (Ghent University) 

An introduction to the Database of Medieval Chinese Texts project: Current functions and future perspectives

The “Database of Medieval Chinese Texts” (DMCT) has been a project ongoing for ca. 10 years at the Department of Languages and Cultures, Ghent University, and its main partner, DILA at Dharma Drum University, Taiwan. Its main aim has been the creation of high-quality digital editions of non-canonical Buddhist texts extant in Dunhuang manuscripts. Originally a by-product of working with the manuscript material, the extraction of variant characters (yitizi 異體字) has developed into a major part of the database project, and the Variants Database module currently contains close to 100.000 character forms. DMCT also contains parts which are not accessible to the public yet, such as the Syntax and Sentence Analysis modules.

In the lecture, I will provide an overview of the structure and functions of the database, and also focus on the way DMCT has “organically” grown in response to specific needs related to research, international collaboration (concretely, importing information from and exporting data to other Database projects), and educational purposes, having developed into an important tool in the infrastructure of our research group, as well as in the training of MA and PhD students at our department.

The last part of the presentation will focus on future perspectives in the development of DMCT.

Dr. Yu-Chun Wang (Dharma Drum Institute of Liberal Arts)

Constructing a Buddhist Knowledge Graph with AI: From Dictionary Concepts to Canonical Relation Triples 

This study presents an AI-assisted workflow for constructing a large-scale Buddhist Knowledge Graph (KG) by transforming dictionary-based conceptual definitions into machine-interpretable relational structures. Buddhist terminology is vast and specialized, making manual entity consolidation and relation extraction extremely labor-intensive. Leveraging recent advances in large language models (LLMs), we develop a multi-stage pipeline that merges synonymous senses, summarizes definitions, classifies doctrinal categories, and extracts subject–predicate–object triples from dictionary entries. Dictionaries provide curated conceptual foundations, allowing LLMs to generate consistent semantic representations suitable for KG construction. This work demonstrates that large-scale KG construction is now feasible through AI, and we are expanding the KG for future public release to support research in Buddhist studies and digital humanities.

Jen-Jou HUNG 洪振洲 (Dharma Drum Institute of Liberal Arts)

From Keyword to Semantics: the Experience of Developing CBETA Corpus Search Functions 

Full-text search is indispensable for researching the complex Buddhist corpus. The Chinese Buddhist Electronic Text Association (CBETA) currently serves over 12,000 queries daily. However, traditional keyword search faces significant challenges, including rigid query formulation, phrasing variations, heterographic variants, and semantic ambiguity. This presentation outlines CBETA’s technical evolution in addressing these challenges. We first review foundational solutions such as variant normalization and the adaptation of the Smith-Waterman algorithm—traditionally used in genomics—for "fuzzy" matching. This enables scholars to locate passages despite scribal errors or version differences. The core focus is the recent integration of Semantic Search using Retrieval-Augmented Generation (RAG). By utilizing vector embeddings, this system moves beyond exact character matching to answer conceptual queries. We present a comparative analysis of our RAG implementations (Beta 1 vs. Beta 2). The optimized architecture, which retrieves top-ranked passages via vector similarity before summarization, successfully reduced response times from 120 to 12 seconds and cut AI token costs by 66%. Finally, we address practical limitations such as hardware costs and hallucinations. We conclude with a roadmap for future developments, specifically the transition towards local LLMs to balance performance with sustainability. These insights aim to provide a valuable reference for the development of next-generation digital archives.

14:50-16:20 各地のデジタル仏教研究とAI活用 

Dr. Patrick McAllister (Austrian Academy of Sciences)

How to Hibernate AND Wake Up Again: Lessons from Open-Ended Editorial Projects

Quite a few projects that produce editions of texts have a neat, linear sequence: following a phase of data collection and curation, a standardised edition is produced. While it is being revised, a workflow to generate outputs — typically an online edition and a printed version — is implemented. Finally, the raw edition with its metadata is put in a long-term archive backed by an institution, and the online edition is kept running for a few years, often at personal cost to researchers.

In this talk, I present some strategies for dealing with projects that do not follow this linear model. These projects typically lack a definitive completion point. My experience draws on projects that aim to produce unbounded text collections. The benefit of such efforts versus the projects focused on a single text is obvious: a uniformly encoded group of texts will be more useful than an isolated edition. But it is equally obvious that this sort of open-ended editorial project poses specific problems. I will argue that three criteria, all organised around the central notion of “utility”, must be in place to provide such projects a good chance of survival: there needs to be an editorial organisation that can survive for long enough to support the project; FAIR principles need to be upheld during and throughout the project, not added at the end only; and the technology underlying the project must remain both functional and open for experimentation. I propose that modern systems of reproducible software stacks can extend the FAIR principles to the software that makes the editions “useful” (or interoperable and reusable) and thereby satisfy the last criterion.

I will discuss two specific cases that illustrate how a project can go through lulls of inactivity and be revived if the organisational, editorial, and technological criteria are met in such a way that they jointly contribute to the project’s utility. One case I will discuss is SARIT, a long-running, loosely organised editorial project that is now waking up from hibernation; the other is VEGEST, a framework for running projects such as SARIT.

• SARIT: https://github.com/sarit/sarit-corpus

• VEGEST: https://gitlab.oeaw.ac.at/vegest/vegest-schema

Bunchird CHAOWARITHREONGLITH, PhD. (Dhammachai Tipiṭaka Project, DCI Center for Buddhist Studies, Thailand)

Digital Tools for Pāli Textual Criticism: Creating a Critical Edition of the Pali Canon from Four Palm-leaf Manuscript Traditions

This talk explores the integration of advanced digital tools into the field of Pāli textual criticism, specifically in the preparation of a critical edition of the Pāli canon. In Thailand, the Dhammachai Tipitaka Project (DTP) utilizes around twenty witnesses—if not more—for each text. These sources include four traditions of palm-leaf manuscripts alongside four printed editions of the Pāli canon: Burmese, Sinhalese, Thai, and European (Be, Ce, Se, and Ee). This is a painstaking task, particularly when incorporating such a high volume of palm-leaf manuscripts. Consequently, advanced digital tools are implemented throughout the entire workflow: manuscript surveying, image segmentation, transcription, text alignment, editing, and apparatus creation. I will demonstrate how we utilize these tools in our editorial work, while also discussing future developments aimed at reducing workload, increasing accuracy, and improving flexibility.

While the DTP has utilized digital tools for over a decade, the limitations of this legacy system—combined with recent advancements in digital and AI technology—have necessitated the creation of a completely new system. Manuscript transcription requires greater accuracy, and text alignment must be more effective when encountering complex passages. Furthermore, the tasks of editing text and preparing the critical apparatus need to be more intuitive. We are integrating these capabilities into our new system and are exploring all possibilities, including AI and OCR. This presentation outlines how we are addressing these challenges.

Sebastian Nehrdich (Tohoku University)

Translation, OCR, and Semantic Retrieval: Current Status and Future Outlook of the Dharmamitra Ecosystem

This presentation gives an overview of the current status and future outlook of Dharmamitra, a collaborative research ecosystem between Tohoku University, the Tsadra Foundation, and the Berkeley AI Research Lab. I will examine how recent advances in Large Language Models facilitate a new paradigm for Buddhist philology and how Dharmamitra explores these through three main technological vectors: First, optical character recognition to digitize classical texts with high accuracy. Second, machine translation systems based on LLM and semantic search technology with high precision for Sanskrit, Pāli, Tibetan, and Chinese. Third, vector-based semantic retrieval to allow seamless search and retrieval across multilingual corpora. I will demonstrate how Dharmamitra can be used to dramatically shorten the time needed for finding relevant passages in large, multilingual corpora, and in what way this might affect textual studies of Buddhist material in the future.