This paper introduces a cosine-based stylometric approach to re-examine the unsigned English editorials and opinion pieces of The Independent (Tongnip Sinmun) from 1896 to 1899, aiming to explore Yun Chi-ho (尹致昊)'s potential English public voice. Although these texts are often associated with Yun, a key bilingual intellectual in late 19th-century Korea, the evidence for this connection remains uncertain. Anglophone research on Korean reformism and sovereignty (Schmid 2002 / Em 2013) situates The Independent within the broader context of Korea’s reformist public sphere but leaves unclear how the individual authorial voice aligns with the paper’s institutional and editorial English texts. The issue isn’t just that many English articles lack signatures. In the bilingual, collaborative journalism of late Chosŏn and early Korean Empire, unsigned English texts could result from multiple stages: drafting, translating, editing, institutional normalization, and newspaper style conventions (Yu 1996 / Chae 2006). Thus, authorship should be understood as a graded stylistic proximity rather than a definitive label, shifting the main question from "Did Yun write this?" to "How does a Yun-related public English voice compare stylistically to his verified private English writings?" By ‘public voice’, this paper means a measurable stylistic profile derived from unsigned newspaper texts, established only after passing proximity and contrastive distinctiveness tests.

The analysis focuses solely on the publication period of The Independent's English edition. Yun's verified diary entries are also limited to those from 1894 to 1900, ensuring comparisons are made within the same historical context. This approach avoids the fallacy of comparing 1896–1899 editorials with Yun's writings from the 1880s or from the post-1910 period, thereby preventing anachronisms. This temporal boundary is considered a strength of the study: it allows for a controlled comparison based on time, rather than using all available materials opportunistically.

The corpus contains limited labels. Some texts are explicitly labeled, such as Yun's verified English diary entries from 1894–1900, Philip Jaisohn(徐載弼)'s signed editorials in The Independent, Yu Kil-chun's English correspondence, contributions by Homer B. Hulbert to The Korea Review (vols. 1–2, 1901–1902), and Horace N. Allen's diplomatic and personal documents. However, unsigned texts from The Independent remain unlabeled. Additionally, the Korean Repository (1892–1898) provides a supplementary reference corpus from Western observers. The analysis utilizes this sparse labeling, using the labeled texts as references to contextualize the unlabeled content.

Each text is represented as a stylometric vector combining character n-grams (n∈{3,4,5}), The analysis includes function words, punctuation patterns, sentence-length percentile bins, and word-length percentile bins, along with the character-start n-gram layer, known for its robustness across genres (Stamatatos 2013). Semantic embeddings like SBERT and SimCSE are intentionally omitted to avoid conflating topical similarity with authorial style—a significant concern given the recurrence of themes such as reform, civilization, education, and public enlightenment across multiple authors. Cosine similarity is consistently used as the distance measure, supported by experimental data showing that vector-length normalization plays a crucial role in improving stylometric distance accuracy (Evert et al. 2017).

The framework reports three dimensionally distinct indicators, hierarchically organized so that each presupposes the previous.

L1 — Stylistic proximity. For each unsigned text xi, stylistic proximity is defined as

Proximityxi=maxd∈DYun1894--1900⁡cos⁡xi,d

where DYun 1894–1900is the temporally aligned authenticated Yun diary corpus. Higher values indicate closer alignment to Yun's authenticated private English style. This first indicator ranks all unsigned texts according to their maximum stylistic proximity to Yun's diary segments.

L2 — Contrastive distinctiveness. The proximity score is meaningful only if Yun is closer to xithan alternative authors are. The central inferential burden of the framework, therefore, falls on L2 rather than L1: a text is treated as Yun-related only when its proximity to Yun survives contrastive comparison with both Western-observer English and Korean-authored English. Inspired by the contrastive logic of Impostors method (Koppel / Winter 2014), contrastive distinctiveness is defined as

Distinctivenessxi=maxd∈DYun1894--1900⁡cos⁡xi,d-maxa∈{Allen,Hulbert,Jaisohn,Yu}⁡maxm∈Ca⁡cos⁡xi,m

where Cais the labeled corpus for author or reference group a. Both terms are maximum points in a set, ensuring that L1 and L2 are measured within the same unit. Positive values denote Yun-specific stylistic alignment rather than general Korean English or Western-observer English. This indicator addresses a key confound: the possibility that proximity to Yun reflects shared topics, nationality, educational background, or bilingual reformist discourse, rather than the author's style. Pipeline performance is validated using known-author comparisons—such as Hulbert and Appenzeller in The Korean Repository—and leave-one-out classification of Jaisohn's signed editorials and Yu Kil-chun's letters. Since the labeled corpus includes epistolary materials that may differ significantly from editorial prose, potentially inflating Yun-specific contrast scores, the analysis also provides a Jaisohn-only contrast as a robustness check. To account for potential outliers affecting point-to-set maxima, max-based distinctiveness is additionally compared with top-k mean distinctiveness as another robustness measure.

L3 — Public shift. Texts that pass both L1 and L2 thresholds — that is, texts both close to Yun and distinguishable from alternative authors — constitute the high-confidence public-voice subset E**. The substantive question is then how this possible Yun-related public voice differs from Yun's private diary baseline:

Public Shift=1-cos⁡VYunprivate,VE**public

This indicator assesses the gap between Yun's verified private English style and the average style of the high-confidence public voice group. It does not serve as conclusive proof of authorship. Instead, it captures a stylistic shift between private and public voices within a carefully selected subset created under strict prior conditions. Therefore, L3 is a conditional metric: it is only computed after L1 and L2 establish a credible Yun-specific subset, and it reflects stylistic displacement rather than standalone evidence of authorship.

Each of the three indicators is displayed both in total and during two historically significant periods: the reformist expansion around the 1896 Agwan p'ach'ŏn (俄館播遷) and the founding of the Independence Club, as well as the later phase marked by Jaisohn's 1898 departure and the newspaper's closure in 1899. Sensitivity to thresholds is demonstrated across L1 thresholds (0.5, 0.6, 0.7) and L2 thresholds (0.05, 0.10, 0.15), creating a nine-cell sensitivity grid. When results remain consistent across thresholds, they strongly support Yun-specific stylistic alignment. Fluctuations are not simply technical issues but may reflect the historical context, specifically the interplay of individual style, editorial influence, and institutional voice. The statistical significance of the public-shift indicator is evaluated using null distributions generated from period-shuffled pairings.

Two outcomes offer equally valuable insights into the historical question. If high-distinctiveness Yun-aligned texts cluster during the early reformist period and fade after Jaisohn's departure, this supports the view of The Independent as a platform for individual reform voices at its founding. Conversely, if Yun-aligned texts continue or grow stronger in later periods, it indicates the institutional strengthening of a Yun-affiliated editorial voice that survived Jaisohn's direct involvement. Both scenarios help clarify how individual styles and institutional voices intersected in late-1890s reformist journalism.

As a falsification test, we apply the same procedure using a non-Yun labeled corpus as the positive seed. If the framework generates equally strong ‘Yun-like’ clusters in this placebo scenario, the result is considered non-discriminative. This test aims to assess whether the framework captures his unique stylistic proximity or simply reflects general genre, topic, or bilingual reformist discourse similarities.

Methodologically, this paper introduces a reproducible, lightweight framework that employs cosine similarity as a single tool across three levels of inference: proximity, distinctiveness, and shift. The hierarchical design directly tackles a common critique of stylometric attribution in small bilingual corpora—that observed proximity may reflect topical, generic, national, or linguistic influences rather than true authorial style. The framework's results are comparable to traditional attribution consensus in two ways: when computational distinctiveness aligns with conventional attribution, it offers quantitative support; and when it diverges, it highlights texts that merit further historiographic review. Substantively, it provides a ranked, contrastively validated list of unsigned texts most likely aligned with Yun's authenticated English voice, along with a quantitative measure of how that voice differs from his private diary style during The Independent's publication period. This approach offers a computational means to reassess Yun Chi-ho's possible English public voice in The Independent. It avoids reducing editorial mediation to mere individual intention and maintains the question of Yun's authorship beyond inherited attributional consensus. More broadly, it shows how authorship-related analysis can be performed on small, anonymous, and collaboratively created historical corpora without forcing uncertain texts into fixed categories.

Keywords: Yun Chi-ho; The Independent; cosine similarity; stylometry; bilingual journalism; digital humanities