In regulated industries, “compliance” in multilingual content is routinely treated as a single, undifferentiated requirement, and automated translation pipelines inherit this monolithic view. This paper, the third stage of a postdoctoral research programme on Knowledge Graph-Mediated Translation (KGMT) at the Ionian University, argues that compliance is a stack of qualitatively distinct constraints, and that distinguishing them is a prerequisite for trustworthy automation and a principled division of labour between machine and human evaluator.
Earlier stages of the programme demonstrated that knowledge graph injection transforms compliance detection in LLM translation of dermocosmetics content (EN→EL, EN→FR) from stochastic to deterministic and auditable, raising detection recall from 24% to 100% for regulatory violations across a dermocosmetics corpus under three pipeline conditions. Building on these results, this paper introduces a taxonomy of seven compliance submodules: legal–regulatory, terminological, genre, stylistic–register, brand, cultural–locale, and accessibility, grounded in register theory, stylistics, genre analysis, legal translation studies, and industry quality frameworks, and ordered along a hard-to-soft constraint gradient. While initially validated in dermocosmetics, the taxonomy is designed as a domain-transferable framework for other regulated industries.
Methodologically, the study is deliberately human-analytical: the taxonomy is validated not through further engineering but through systematic re-annotation of existing experimental outputs across the three pipeline conditions, using a double-annotation and adjudication protocol reporting inter-annotator agreement and error distributions. The contested boundary between brand and register compliance serves as the critical test of the taxonomy’s robustness.
The findings speak directly to the question of the human agent in technology: while hard constraints are demonstrably automatable, the soft submodules (e.g. register, brand voice, cultural appropriateness) prove to be where human compliance judgement resists formalization. The taxonomy thus functions as an implementation blueprint for future knowledge graph extension and an empirical map of where, in compliance-critical translation, the human evaluator remains irreplaceable.
Keywords
compliance, knowledge graphs, machine translation, annotation, stylistics, error taxonomy, human-in-the-loop
Dr. Viveta Gene is a Postdoctoral Researcher at the Ionian University (Department of Foreign Languages, Translation and Interpreting), in collaboration with Assoc. Prof. Vilelmini Sosoni, where her work centers on Knowledge Graph-Mediated Translation (KGMT) as a framework for compliance-aware multilingual communication in regulated industries. She holds a PhD in Translation and New Technologies from the same institution. A recognized industry expert with over 20 years of experience, Viveta serves as Head of Global Localization Solutions and Innovation at Intertranslations UK, where she leads corporate AI localization strategies and workflows. She sits on the Board of Directors of the Globalization and Localization Association (GALA) and is the author of the official GALA Common Machine Translation Post-Editing Protocol.
Vilelmini Sosoni holds a PhD in Translation and Text Linguistics from the University of Surrey and is Associate Professor at the Department of Foreign Languages, Translation and Interpreting at the Ionian University where she also serves as Associate Head. She is the Director of the Research Laboratory for Specialised Translation and Language Technologies (STraLaT) and a member of the University’s Gender Equality and Anti-Discrimination Committee. She also has extensive professional experience having worked as a professional translator, editor, content creator and audiovisual services provider. She has participated in several EU-funded projects and has published extensively in the areas of the Translation of Institutional Texts, Machine Translation and Audiovisual Translation and Accessibility, including Audio Description and SDH.
