The Dainihon Shiryō is a foundational Japanese historical source collection, structured as a chronological account of events. The Dainihon Shiryō Integrated Database (DNS-IDB) builds on this by representing each event as an individual unit, linked to place names, personal names, temporal information, source citations, and supporting textual excerpts. While this metadata enables event-centric search and indexing, it does not yet sufficiently represent the relationships among events—such as causality, continuity, or response—that are central to historical interpretation. In this paper, we propose a method for constructing a knowledge graph from DNS-IDB to visualize and infer inter-event relationships (e.g., “Event A prompted Event B”, “Event C followed Event D in the same administrative chain”, “Event E was a response to Event F”) and thereby enhance historiographical insight into Japanese history.

Our workflow consists of the following stages. First, event data from DNS-IDB is pre-processed: place names and personal names are normalized (using a custom authority list of historical Japanese toponyms and persons), and event summaries and source-text excerpts are extracted. Second, we represent each event as a vector embedding using a multilingual Sentence-BERT model applied to the event summary and supporting excerpt. This enables measurement of semantic similarity among events in terms of actors, places, and themes. Third, candidate relations among events are inferred using a hybrid rule-based and machine-learning approach: rule-based patterns (e.g., “appointment → dispatch → report”, “conflict → treaty → reorganization”) are combined with a classifier trained on annotated event-event pairs to detect causality, temporal succession, or administrative dependency. We further impose temporal and spatial constraints (e.g., events must be in a plausible sequence, within similar regions) to reduce spurious links. Fourth, the resulting graph—nodes representing events, persons, places, and sources; edges representing relations such as causedBy, followedBy, and responseTo—is implemented in a Neo4j backend and visualized in interactive dashboards. Examples of these relationships are shown in Figure 1. Scholars can explore event-chains, trace administrative decision-pathways across regions, or zoom into micro-histories of local incidents.

Figure 1: Examples of visualizing historical events using a knowledge graph Figure 1: Examples of visualizing historical events using a knowledge graph

Related Research:
Knowledge graphs (KGs) have gained traction in digital humanities to move beyond static entity-centric representations toward dynamic, interconnected knowledge structures. For example, EventKG (Gottschalk & Demidova 2018) produced a multilingual event-centric temporal KG containing hundreds of thousands of events and millions of temporal relations. Hassanzadeh (2021) built a KG of events and consequences drawn from Wikipedia and Wikidata to explore forecasting and consequence-analysis of events. In the cultural-heritage domain, Morvillo & Mecella (2024) examine the challenges of integrating multiple humanities KGs, and Schögl et al. (2025) describe the InTaVia knowledge graph, which models biographical, object- and event-centric data using CIDOC-CRM. On the extraction side, Liu et al. (2025) propose an Event Causality Graph method to detect cross-document causal relations between events. These studies provide strong precedents for our methodology of embedding-based similarity, diagnostics of causation, and graph-based visualization. Yet to our knowledge, none apply this approach explicitly to a large Japanese historical corpus such as DNS-IDB, nor focus on the editorially defined event-unit as the primary node. Our project fills that gap by treating each DNS event as a first-class node and by modelling inter-event relations that reflect Japanese archival historiography. Knowledge graphs (KGs) have gained traction in digital humanities to move beyond static entity-centric representations toward dynamic, interconnected knowledge structures. For example, EventKG (Gottschalk & Demidova 2018) produced a multilingual event-centric temporal KG containing hundreds of thousands of events and millions of temporal relations. Hassanzadeh (2021) built a KG of events and consequences drawn from Wikipedia and Wikidata to explore forecasting and consequence-analysis of events. In the cultural-heritage domain, Morvillo & Mecella (2024) examine the challenges of integrating multiple humanities KGs, and Schögl et al. (2025) describe the InTaVia knowledge graph, which models biographical, object- and event-centric data using CIDOC-CRM. On the extraction side, Liu et al. (2025) propose an Event Causality Graph method to detect cross-document causal relations between events. These studies provide strong precedents for our methodology of embedding-based similarity, diagnostics of causation, and graph-based visualization. Yet to our knowledge, none apply this approach explicitly to a large Japanese historical corpus such as DNS-IDB, nor focus on the editorially defined event-unit as the primary node. Our project fills that gap by treating each DNS event as a first-class node and by modelling inter-event relations that reflect Japanese archival historiography.

Contribution:
This paper contributes in three major ways: (1) It presents a workflow for transforming a Japanese event-centric historical database into a machine-readable knowledge graph capable of relation inference; (2) It bridges TEI/metadata processing, embedding computations, rule- and ML-based relation extraction, and graph visualization into a reproducible pipeline; (3) It opens up the possibility of integrating DNS-IDB with external data infrastructures and enables historians to explore causal and regional dynamics at scale. Preliminary experiments over a subset of DNS events show that our embeddings capture latent similarity clusters of events across decades and regions, and the inferred relations align with known administrative chains (e.g., appointment → dispatch → report) and regional conflict sequences. In doing so, we hope to advance computational historiography by providing new tools for exploring causality, continuity and spatial-temporal dynamics in Japan’s richly documented historical record. This paper contributes in three major ways: (1) It presents a workflow for transforming a Japanese event-centric historical database into a machine-readable knowledge graph capable of relation inference; (2) It bridges TEI/metadata processing, embedding computations, rule- and ML-based relation extraction, and graph visualization into a reproducible pipeline; (3) It opens up the possibility of integrating DNS-IDB with external data infrastructures and enables historians to explore causal and regional dynamics at scale. Preliminary experiments over a subset of DNS events show that our embeddings capture latent similarity clusters of events across decades and regions, and the inferred relations align with known administrative chains (e.g., appointment → dispatch → report) and regional conflict sequences. In doing so, we hope to advance computational historiography by providing new tools for exploring causality, continuity and spatial-temporal dynamics in Japan’s richly documented historical record.

Acknowledgment:
This work was supported by JSPS KAKENHI Grant Numbers 24H00011, 25K24614, 25H01239, 25H00482.