Daejeon, July 27–31
What does a DH project look like that engages with historical cross-cultural social connections using AI and promotes data diversity and equity? The paper attempts to answer this question by presenting a prototype of LLM-assisted knowledge graph creation for a small-scale feminist DH project about transpacific networks of women writers and artists in the first half of the twentieth century. The project addresses enduring Eurocentrism and the skew toward the Atlantic over the Pacific in scholarship on both women’s internationalism and feminist digital projects, thereby recovering a largely forgotten feminist network across the East-West divide.
Feminist scholars have problematized the neglect of women’s contributions to internationalism and documented historical cases of women’s collective efforts (Rupp 1997; Sluga and Clavin 2017; Owens and Rietzler 2021; Paisley 2009; Yasutake 2024). Although significant, these studies have focused on large organizations with headquarters in Europe or the United States. This project, in contrast, explores women’s grassroots cross-cultural networking centered in Tokyo.
Knowledge graphs have three advantages over other data models for carrying out this project. Firstly, their scalability, modifiability, and tolerance for sparse data lend themselves better than relational databases to ongoing projects that require incremental data population from diverse textual sources, including newspaper articles, diaries, correspondence, and event programs, as well as visual sources such as photographs. Secondly, graph-based models, which store relationships explicitly as first-class elements and can link to external datasets, facilitate more nuanced social network analysis (Pan 2022, 6). They enable us to closely trace developing interconnected networks of diverse agents. Thirdly, knowledge graphs improve the findability and accessibility of data when linked to external databases. The scarcity of research on international women’s clubs in Tokyo is partly due to the obscurity of institutional records: more established organizations have greater online presence and provide researchers with catalogues and easier access to documents. This knowledge graph project will enhance the discoverability of these overlooked sources through linked data, thus promoting productive connections between feminist databases and the data diversity and nuance that the humanities bring to data science (Smith Elford and Meagher 2023, 5; Nurmikko-Fuller 2024, 104).
This prototype focused on the Tokyo Penwomen Club, an international women’s press club active in Tokyo from 1926 to approximately 1931. The dataset comprised 84 English-language newspaper articles from the Japan Times Archives, supplemented by biographies and an event program. We applied the three-step Extract–Define–Canonicalize (EDC) framework (Zhang and Soh 2024) to OCR texts, with three principal modifications for humanities research: source-grounded extraction, frame-aware relation extraction, and human-curated canonicalization.
The extraction stage comprised Named Entity Recognition (NER) across five entity classes—persons, organizations, works, places, and events—followed by relation extraction (RE), with both steps iterated through close reading. Every extracted triple was linked to its source passage, keeping interpretive decisions traceable to documentary evidence. RE employed a parallel two-track approach. The first track applied direct RE, producing triples with natural-language predicates and explicit object-class typing. The second track used a frame-aware pipeline that began with a summarization step capturing the text’s macroscopic discourse structure—who is speaking, about whom, and in what capacity—before NER and RE. Outputs were merged and deduplicated. While EDC’s canonicalization stage relies on LLM-generated definitions matched via vector similarity, we used human-curated authority tables instead: one linking entities to canonical names, aliases, and Wikidata identifiers, the other mapping relation phrases to a canonical predicate schema. Both tables were swappable to accommodate different interpretive profiles. Analogous to EDC’s refinement phase (EDC+R), which feeds prior extractions back as contextual hints, our human-curated tables were reinjected into the normalization step so that later extraction passes benefited from expert corrections.
Against a manually annotated gold standard of 225 entities across seven files, NER yielded precision of 0.56 and recall of 0.60. These figures are indicative rather than definitive: many apparent false positives are legitimate extractions the annotation did not capture, as the annotator frequently encoded descriptive phrases as single entities while the pipeline resolved them into relational triples. Given this difficulty, the precision of the 259 relations extracted from the sample files was evaluated by direct inspection at 94.2%, though some triples were borderline and residual near-duplicates from the two-track merge were not fully accounted for.
Performance varied substantially by text type: a structured event program yielded NER F1 ≈ 0.77, while discursive prose reached F1 ≈ 0.36; the picture reversed for RE, where errors concentrated in dense nominal lists rather than prose. The asymmetry reflects the contrasting demands of the two tasks—lists provide clean entity mentions, while prose furnishes the contextual elaboration that grounds relational inference.
Despite the small corpus, we made two findings and observed accompanying challenges. Firstly, most regular members were inconsistently identified by their own names or their husbands’ names across papers, whereas Yei Theodora Ozaki was consistently referred to as “Madame Ozaki” or “Madame Yukio Ozaki.” This name variation poses canonicalization challenges common to feminist DH projects (Booth 2015, 88). Secondly, members’ countries of affiliation, such as nationality, citizenship, and residence, were fairly diverse, although the USA and Japan constituted the majority (Figure 1). We settled on the ambiguous and inclusive term “affiliation” here because of historical discrimination against women in nationality law (Rubio-Marin 2022, 178; Knop 2001, 96–99). EDC’s bottom-up approach and our use of multiple schemas allow us to handle these messy findings without reduction and simplification. This responds to LOD researchers’ call to embrace data ambiguity and apply data-driven ontological structures that enable accurate knowledge representation and nuanced interpretation (Brown 2022; Nurmikko-Fuller 2024, 17, 100).
Figure 1. Countries of Affiliation
If humanities scholars with limited budgets can construct expert knowledge graphs using AI and link them up, this constellation of small-scale datasets will accelerate our research and raise new questions. It will also improve data reusability and sustainability since research documents and findings will be more readily shared. This linked humanities ecosystem is what our project aspires to.