Hierarchical Flattening, Neural Machine Translation, and the Organizational Semiotics of Korean Honorifics in Workplace Drama Subtitles: A Work in Progress

Introduction

The organizational adoption of AI-driven translation tools for cross-cultural professional communication has prompted renewed attention to questions about what kinds of meaning computational systems can and cannot reliably carry across languages. This paper takes up one such question through the analysis of Korean workplace drama subtitles, examining how neural machine translation (NMT) systems handle the grammatically encoded hierarchical information that structures Korean professional discourse. Korean dramas such as Misaeng (2014), Start-Up (2020), and Itaewon Class (2020) have circulated widely across international platforms, and their reception by global audiences depends substantially on subtitle translation. These texts are organized around Korea’s honorific system, in which the contrast between formal registers (jondaemal) and informal registers (banmal) encodes socially consequential information about organizational positioning. When a junior employee addresses a superior using honorific verb morphology, Korean-competent audiences read that choice as an unambiguous index of institutional hierarchy. This paper examines what happens to that relational information when AI translation systems render such exchanges for Anglophone audiences, and puts forward the term hierarchical flattening to describe the systematic suppression of hierarchical meaning observed in the materials analyzed. The paper draws on an initial analysis of five dialogue exchanges from a corpus of fifteen planned exchanges and presents early findings alongside a developing methodological framework.

Situating the Research

Research in organizational communication has established that hierarchy shapes the structure and texture of workplace interaction, with messages encoding relational positioning through forms of address, politeness strategies, and register choices (Jablin / Putnam 2001). Cross-cultural management scholarship has further demonstrated that societies characterized by high power distance, among which Korea is consistently placed, tend to institutionalize such encoding linguistically, while low power distance societies such as the United States characteristically minimize explicit hierarchical marking in professional discourse (Hofstede 2001). The implications of this contrast for translation practice are considerable. Translation scholars working in pragmatic and functionalist traditions have argued that the translation of professional communication requires not merely the transfer of propositional content but the preservation of relational information about the positioning of interlocutors (Venuti 2008). Screen translation research has similarly examined how subtitle translation negotiates cultural meaning for audiences operating within different semiotic economies (Diaz Cintas / Remael 2007). The organizational and linguistic dimensions of this problem are, in the existing literature, not yet fully integrated with questions about what NMT systems specifically introduce when handling languages whose grammatical architecture embeds hierarchy in morphosyntactic categories that English does not share.

The computational dimension has a parallel scholarly trajectory. Transformer-based NMT has achieved widespread deployment across professional and consumer contexts since its foundational articulation by Vaswani et al. (2017), but evaluation practices in the field have tended to privilege surface accuracy metrics such as BLEU scores over measures of cultural or pragmatic appropriateness. Korean honorifics have been identified as a domain of particular difficulty for automated systems (Lee et al. 2021), and more recent scholarship has begun to address cultural adaptation as a structural challenge for cross-lingual natural language processing (Hershcovich et al. 2022). Management research, while attentive to the ways in which language barriers complicate trust formation and knowledge transfer in multinational teams (Tenzer et al. 2014), has not yet examined how AI translation tools may introduce systematic and directional information loss of the kind this paper begins to investigate. The present work is situated at the intersection of these bodies of literature, at a moment when the DH2026 subtheme of “Translating | Translinguality in the Era of AI” makes such questions particularly pertinent.

Methodology

The study employs a comparative corpus-analytic framework applied to fifteen dialogue exchanges drawn from three Korean workplace dramas produced between 2020 and 2024. The three dramas were selected to represent distinct organizational settings, specifically technology startups, traditional corporate environments, and service industries, on the grounds that honorific usage patterns, while grammatically systematic, may carry varying social force across these contexts. Individual exchanges were selected to cover organizational scenarios of cross-cutting significance: superior-subordinate address, peer collaboration, client-facing communication, and leadership under pressure. Each exchange contains clearly marked formal or informal speech, and the corpus as a whole is designed to cover the principal honorific registers operative in Korean professional discourse. Exchanges range from four to twelve dialogue turns.

All source materials are publicly available. Korean-language subtitle files were accessed through open online archives. Translation corpus consist of professional Netflix subtitles and/or community Viki translations. The identification of honorific marking in the Korean source proceeds through close analysis of verb morphology and speech-level markers, supplemented by the author’s sustained engagement with Korean workplace media and foundational language training. A methodologically notable additional source of verification is provided by the public comment sections of Viki discussions, where Korean-competent viewers routinely flag translation errors and honorific mistranslations. This community-generated metalinguistic commentary, produced organically by native speakers in response to specific translation choices, constitutes a transparent and reproducible form of error annotation that the study treats as a secondary verification layer.

Each exchange is coded along four analytically distinct dimensions. Hierarchy preservation examines whether speech-level distinctions grammatically present in the Korean source are carried into the target text. Communicative strategy examines the specific linguistic resources deployed, including vocabulary selection, pronoun use, syntactic subordination, role-marking conventions, and hedging practices. Organizational accuracy asks whether a target-language reader would correctly infer the relational positioning of the interlocutors from the translated exchange. Managerial implication examines what the translation signals about the institutional role and status of the speaker. These dimensions were developed from the translation studies and cross-cultural communication literature and operationalized through a pilot coding round prior to application to the full dataset. Coding is currently complete for five of the fifteen planned exchanges.

Preliminary Findings

Analysis of the five coded exchanges reveals a consistent pattern with theoretical implications. Both AI systems under examination, specifically Google Translate and ChatGPT-4, render formal superior-subordinate exchanges and informal peer exchanges into a functionally indistinguishable neutral English register, producing the hierarchical flattening the paper identifies. Across the five exchanges, approximately sixty percent of machine translations did not preserve hierarchical distinctions that are grammatically explicit in the Korean source.

The contrast with human translation is instructive. In one exchange involving a deferential approval request from a junior employee to a superior, Google Translate produced “Can you approve this proposal?”, a formulation that, by the conventions of American professional communication, implies peer-level address rather than institutional deference. ChatGPT-4 produced “Could you review and approve this proposal?”, which is marginally more formal in register but not substantively more hierarchically marked. The corresponding Netflix subtitle, by contrast, deployed multiple hedging and deference markers in combination, producing a rendering in which the status differential remains legible to Anglophone readers without requiring knowledge of Korean grammatical conventions. The Viki community translation addressed the same differential through explicit role-title marking, a strategy involving different trade-offs but similarly preserving organizational information.

Initial analysis further suggests a systematic difference in the translational repertoire available to AI systems compared with human translators. AI systems appear to rely predominantly on vocabulary-level formality cues, such as the use of “please” or the conditional “could”, while human translators draw on a wider set of resources including pronoun selection, syntactic subordination, appositive role-naming, and pragmatic hedging. A supplementary finding of interpretive interest is that when ChatGPT-4 was tested with prompts that made the hierarchical relationship between interlocutors explicit in the input, translation quality improved substantially. This finding suggests that the system possesses some capacity for honorific-sensitive translation but does not reliably activate that capacity on the basis of grammatical markers alone, indicating that the difficulty may be partly one of context inference rather than purely of architectural incapacity. These findings are preliminary and require validation through the complete fifteen-exchange corpus. Future work will examine whether the observed patterns vary by organizational context, drama genre, or the nature of the hierarchical relationship encoded, and will consider how findings may generalize to typologically comparable honorific languages.

Implications and Conclusion

This paper identifies a form of information loss that operates at the level of relational and hierarchical meaning and that current NMT evaluation frameworks are not designed to detect. The dominance of English and of low power distance communicative norms in the training data of large language models may contribute to a form of cultural homogenization whose consequences extend beyond entertainment subtitle translation into the organizational deployment of AI translation tools in international professional settings, though this broader claim warrants further empirical investigation. The aligned multilingual corpus of honorific-marked workplace dialogue constructed for this project will be made publicly available upon completion, providing a reusable resource for both translation studies and computational research on honorific systems. More broadly, the paper raises questions about the politics of AI-mediated global communication: whose organizational knowledge is preserved, and whose may be systematically attenuated, when the computational infrastructures of cross-cultural communication reflect a narrow subset of the world’s communicative norms. Understanding these dynamics, even in their early and provisional dimensions, is relevant to the responsible deployment of AI in international settings, whether in entertainment translation, scholarly communication, or organizational practice.