Research — Aron van de Pol

Journal articles

Published

Interpretable Visual Feature Discovery using Multiple Instance Learning: A Case Study of Colonial Korean Print

Aron van de Pol, Jelena Prokić, Angus Mol, 2026, Computational Humanities Research

Journal Code

This paper explores how neural networks, particularly Multiple Instance Learning (MIL), can advance computational analysis of historical visual materials. While Distant Viewing (Arnold and Tilton, 2023a) has shown how computer vision can analyze visual materials at scale, interpreting how these models identify distinctive features remains a key methodological challenge. Although deep learning approaches achieve high classification accuracy, understanding which visual elements inform their decisions is often unclear, limiting their utility for humanities research where interpretability is crucial. We address this challenge through an innovative application of Multiple Instance Learning to historical print analysis. Using colonial Korean printshops (1910–1945) as our case study, we demonstrate how MIL’s interpretable architecture can reveal distinctive visual features while maintaining classification accuracy (Cai et al., 2024). This approach allows us to examine not just what distinguishes different printshops’ outputs, but specifically which typographic elements and patterns the model uses to make these distinctions. Drawing on a large digitized corpus from four major colonial Korean printshops, we demonstrate how MIL, originally developed for medical imaging analysis, can effectively classify and help interpret historical visual materials. The model achieves 92% accuracy while providing interpretable attention maps that highlight historically meaningful features. The results demonstrate how computational methods can reveal subtle but consistent variations in common textual elements, such as the rendering of ieung (ㅇ) in frequently occurring grammatical particles (josa 조사) or the bottom of the tigut (ㄷ). This methodological contribution extends beyond Korean print history, offering a framework for interpretable computational analysis of visual materials more broadly.
Forthcoming

Buying Type, Selling Print: Type Foundries, Print Shops, and the Commercial World of Colonial Korean Printing, 1911–1945

Aron van de Pol, 2026, Seoul Journal of Korean Studies

This article draws on approximately 2,350 advertisements placed by print shops, type foundries, and in-house printing departments in three major colonial Korean newspapers (Maeil sinbo, Tonga ilbo, Chosŏn ilbo) between 1911 and 1949, extracted from a database of over 623,000 classified advertisements. It identifies roughly 1,150 printing-related entities: 881 independent print shops, 106 incorporated firms, 88 type foundries, 33 in-house departments, and 42 supplemental advertisers. The breakdown documents the commercial structure of an industry whose ordinary work (business cards, government forms, certificates) has been invisible in a historiography focused on literary and political print culture. The type supply chain was dominated by a compact group of Japanese-owned foundries in Kyŏngsŏng, while the downstream printing sector was overwhelmingly Korean-owned (97 percent of identifiable firms). The advertising record fills a gap between existing macro-level arguments about typographic hegemony and micro-level studies of individual printshops, showing that the commercial base of colonial Korean printing was far larger and more geographically dispersed than government statistics have indicated.