Daejeon, July 27–31
With the maturation of Digital Humanities (DH) methodologies, Buddhist textual studies are shifting from digitization toward datafication and semantic modeling. Over the past decades, projects such as the Chinese Buddhist Electronic Text Association (CBETA) have successfully enabled large-scale preservation and full-text access to Buddhist canons. However, keyword-based retrieval remains insufficient for addressing the conceptual density of Buddhist literature, which is characterized by extensive technical terminology, multi-layered doctrinal systems, and pervasive semantic polysemy resulting from Chinese translations.
These characteristics frequently lead to information overload and semantic fragmentation in textual search, limiting scholars’ ability to trace conceptual relations across texts. Knowledge graphs (KGs), which represent knowledge as entity–relation–entity triples, provide a viable solution by transforming unstructured textual materials into machine-interpretable semantic networks. Such representations enable cross-textual reasoning and have become central to graph-enhanced retrieval-augmented generation (GraphRAG) and other generative AI applications.
This study proposes an AI-assisted framework for constructing a Buddhist knowledge graph grounded in authoritative Buddhist dictionaries. By leveraging large language models (LLMs) to extract structured doctrinal knowledge, the project aims to reduce hallucination in domain-specific AI applications and to establish a computable semantic infrastructure for Buddhist studies within the DH paradigm.
Existing research on Buddhist knowledge graph construction can be broadly categorized into two major technological trajectories. The first emphasizes large-scale entity extraction from general-purpose web resources. A representative example is the KG-Buddhism project proposed by Wu et al. (2017), which constructed a Chinese Buddhist knowledge graph using semi-structured data from online encyclopedias such as Baidu Baike and Chinese Wikipedia. While this approach demonstrates the technical feasibility of extracting Buddhist-related entities at scale, its reliance on popular knowledge repositories limits its suitability for scholarly research. Such sources lack the terminological rigor and doctrinal precision characteristic of authoritative Buddhist dictionaries, resulting in knowledge representations that often flatten or oversimplify complex doctrinal relationships.
The second trajectory focuses on library science and linked open data (LOD) approaches. Roux (2023), in his account of the Buddhist Digital Resource Center (BDRC), documents the migration of bibliographic records from relational databases to RDF-based infrastructures using standards such as BIBFRAME and IIIF. This approach excels at modeling textual transmission, versioning, and cataloging metadata, enabling high levels of interoperability across digital collections. However, its primary concern lies in bibliographic metadata rather than in the internal semantic structure of Buddhist doctrinal content.
In the Taiwanese context, earlier digital Buddhist studies concentrated on lexical extraction techniques, such as suffix arrays and PAT-trees for term discovery (Chien, 1997). Subsequent work by Tang and Hsieh (2012) explored Buddhist lexical ontologies based on CBETA texts, drawing on OntoLex and WordNet-like taxonomies. Although these efforts structured terminological hierarchies, they offered limited means to represent deeper semantic relations among doctrinal concepts.
Recent initiatives, such as the Dharma Setu framework (Lindsey, 2025), have begun exploring the integration of AI and vector space models to connect traditional knowledge systems with modern computational infrastructures. Building upon these developments, our study proposes a distinct methodological pathway: it places authoritative Buddhist dictionaries at the center of knowledge graph construction and employs LLMs for deep semantic extraction. In doing so, it bridges the gap between encyclopedic approaches and bibliographic infrastructures by focusing explicitly on the logical structure of Buddhist doctrine.
The study pursues three interrelated objectives. First, it aims to achieve conceptual systematization and standardization by integrating multiple authoritative Buddhist dictionaries, such as the Foguang Buddhist Dictionary and Ding Fubao’s Dictionary of Buddhist Studies. Through the consolidation of synonymous and variant terminologies, the approach addresses long-standing issues of terminological inconsistency and transforms unstructured dictionary definitions into normalized semantic representations.
Second, the study seeks to develop an automated pipeline for extracting subject–predicate–object (S–P–O) relation triples using LLMs. Emphasis is placed on capturing doctrinal logic rather than surface-level lexical associations, enabling functional, attributive, and classificatory relationships to be formalized as canonical triples.
Third, the study aims to support advanced AI-driven applications by constructing a knowledge graph endowed with semantic reasoning capabilities. The resulting resource is intended to enhance semantic search, intelligent question answering, and graph-enhanced generation, thereby improving the accuracy and interpretability of AI-assisted research in Buddhist studies.
The proposed framework adopts a five-stage human-in-the-loop workflow that integrates expert oversight with LLM-based semantic processing. In the first stage, word sense consolidation is performed by comparing definitions from multiple dictionary sources to merge synonymous expressions and disambiguate polysemous terms. In the second stage, consolidated definitions are rewritten into standardized semantic summaries, removing philological commentary and extraneous references to produce clean textual inputs.
The third stage introduces hierarchical structure through semantic classification. Thirteen core semantic categories are defined, and an ensemble voting strategy is employed using GPT-4.1, Claude 3.7, and Gemini 2.5. Final classifications are determined by majority voting, reducing model-specific bias and stochastic error.
In the fourth stage, a predefined relation schema is established, encompassing functional, attributive, doctrinal, and textual relations. This schema constrains relation extraction and ensures semantic consistency. Finally, in the fifth stage, LLMs extract structured S–P–O triples in JSON format from the standardized summaries, forming the core of the Buddhist knowledge graph.
To evaluate the effectiveness of the proposed workflow, a quantitative assessment was conducted focusing on the semantic classification stage. A representative sample of dictionary entries was classified using both single-model, zero-shot approaches and the ensemble voting strategy, with judgments by Buddhist studies experts serving as the ground truth. The ensemble approach achieved an agreement rate of approximately 94.9% with expert classifications. The results demonstrate the reliability of LLMs in handling complex doctrinal concepts, the effectiveness of ensemble voting in correcting sporadic errors, and the value of standardized summaries in improving the structural quality of extracted relation triples.
Future work will scale the workflow to full dictionary coverage, potentially encompassing tens of thousands of entities and hundreds of thousands of relations. Human-in-the-loop validation interfaces will be developed to facilitate expert review and iterative refinement. The resulting knowledge graph will be publicly released in RDF and JSON-LD formats to promote reuse and interoperability. Application-oriented development, including knowledge-graph-based question answering systems, will further evaluate the reasoning capabilities enabled by the constructed semantic infrastructure.