Morphological Complexity: The Structural Mismatch No Fine-Tuning Can Fix
To understand why the problem is architectural rather than merely a matter of data volume, it helps to begin with morphology. English is a morphologically shallow language. The word library is a single, essentially atomic unit; possession is expressed by adding separate words. Many European minority languages, like Welsh or Finnish, are morphologically rich: a single written word can encode root meaning, grammatical gender, case, number, and pronominal attachment simultaneously. This morphological density means that each distinct word form counts as a separate vocabulary entry in most standard tokenisation schemes.
The consequence is that models built around English morphological assumptions need to be substantially retooled to achieve comparable coverage of morphologically complex languages. Memory requirements increase, inference slows, and training complexity grows. More fundamentally, a model fine-tuned for Welsh or Irish on top of an English-optimised base architecture is working within vocabulary and structural assumptions that were never designed for those languages. As Isabelle Augenstein, professor of computer science at the University of Copenhagen and a leading European voice on multilingual NLP, has noted in her research on cross-lingual transfer, architectural mismatches between source and target languages systematically limit how well transfer learning can work, regardless of the volume of fine-tuning data applied.
Writing directionality is a related but distinct complication. Models trained on left-to-right text learn directional patterns about how information flows through a sentence. Bidirectional attention mechanisms in modern transformer architectures mitigate this to some degree, but the problem resurfaces in downstream tasks like machine translation and information extraction where word order is critical. For European healthcare applications, this matters: a model that misreads the order of symptoms, dosages, or contraindications in a clinical note is not merely less useful; it is potentially dangerous.
Content Bias: Western Entities Dominate, Even in Local Contexts
Beyond morphology lies a subtler and arguably more insidious problem: content bias. When models are trained on English-dominant corpora, they learn statistical patterns that reflect the cultural and institutional world those corpora describe. Western companies, Western politicians, Western clinical guidelines, and Western drug formularies are vastly overrepresented in digitised text. When these models are then deployed in other linguistic contexts, they reproduce Western patterns as if they were universal norms.
The numbers from multilingual NLP benchmarks are stark. Models recognise Western-origin entities mentioned in non-English text with accuracy scores roughly 27 percentage points higher than equivalent locally-relevant entities. This is not the model being deliberately biased; it is faithfully learning from training data where Western entities appear far more frequently. But for a Welsh-speaking patient using an AI-assisted triage tool, or a Basque-speaking elderly person interacting with a digital health assistant, this bias translates directly into worse outcomes.
Consider a practical healthcare illustration. An AI clinical decision support tool trained predominantly on English-language medical literature will have learned drug names, dosage conventions, and diagnostic coding systems common in US and UK English-language practice. It may handle NICE guidelines competently in English while struggling to apply equivalent guidance from the European Medicines Agency when queries are posed in a lower-resource European language. It is not failing because the model is incapable; it is failing because its training data reflects a particular institutional world. The European Medicines Agency has itself flagged in its 2023 reflection paper on AI in medicines regulation that data representativeness, including linguistic representativeness, is a foundational requirement for trustworthy AI in healthcare, yet it remains systematically unaddressed by the major model providers.
If content bias is insidious because it hides in plain sight, the performance gap between formal and colloquial language registers is stark and measurable. Research on multilingual models consistently shows that accuracy on formal, written, standardised language is substantially higher than accuracy on colloquial or dialectal speech. In benchmarks on morphologically complex languages, the gap between performance on formal written registers and informal spoken-style text can exceed 40 percentage points. That is not gradual degradation; it is a cliff edge.
This matters enormously for healthcare. Patients describing symptoms do not speak in formal registers. They use colloquial vocabulary, regional expressions, and informal constructions. A clinical AI that performs acceptably on formal medical records but fails on informal patient-reported text is not fit for purpose in any real clinical environment. Worse, the populations most likely to use informal registers, people with lower formal educational attainment, elderly patients, people from rural communities, are precisely the populations that most need accessible AI-assisted health services. The technical inequality maps directly onto social inequality: those with the highest formal literacy get workable systems; everyone else gets a tool that fails them regularly.
The EU AI Act, which entered into force in August 2024, classifies AI systems used in healthcare as high-risk under Annex III. High-risk systems are subject to requirements including accuracy, robustness, and non-discrimination across foreseeable conditions of use. Linguistic register variation is unambiguously a foreseeable condition of use in any patient-facing application. Dragoș Tudorache, the Romanian MEP who co-led the European Parliament's negotiations on the AI Act, has stated publicly that the Act's non-discrimination provisions were intended to cover performance disparities rooted in training data as well as algorithmic design. Whether national market surveillance authorities will enforce this interpretation rigorously remains to be seen, but the legal framework is there.
The Extractive QA Gap and Its Clinical Implications
Extractive question-answering, where a model finds and returns a relevant passage from a document in response to a query, is a capability that underpins a wide range of clinical AI applications: summarising patient records, surfacing relevant treatment guidelines, answering clinician queries from electronic health record data. Benchmark performance for English extractive QA sits around 85% F1 (a combined measure of precision and recall). For lower-resource European languages, performance drops to roughly 70% F1, a 15-percentage-point gap that reflects multiple compounding factors.
First, annotated QA datasets for lower-resource languages are orders of magnitude smaller than English equivalents like SQuAD. Models trained on millions of English question-and-answer pairs will outperform models trained on tens of thousands of equivalent pairs in another language, all else being equal. Second, question formation in morphologically complex languages involves structural patterns, such as varied word order and relative clause constructions, that require more training examples to learn reliably. Third, the communities building datasets for minority European languages often lack the institutional resources and annotation infrastructure available to major English-language research groups, meaning dataset quality can be uneven.
The cumulative effect in a healthcare setting is straightforward: fewer patients querying AI tools in minority or regional languages get accurate, complete answers. Given that extractive QA is a building block for clinical decision support, drug interaction checking, and patient-facing symptom checkers, a 15-point performance gap is not a minor inconvenience. It is a systematic reduction in care quality for linguistic minorities.
What European Deployment Actually Requires
The standard industry response to these problems is fine-tuning: take an English-optimised base model, train it on additional data in the target language, and hope the performance gap narrows. Fine-tuning helps at the margins. It does not resolve architectural mismatches. It does not address the content bias embedded in a model's learned world-knowledge. And it does not bridge the formal-colloquial gap if the fine-tuning data is itself predominantly formal.
What genuinely multilingual AI for European healthcare deployment requires is a different set of commitments. Tokenisation schemes must be designed to respect the morphological structures of target languages rather than imposing English-style morpheme boundaries. Training corpora must include colloquial and dialectal text from the start, not as an afterthought. Entity recognition systems must be trained on locally-relevant institutional knowledge, including European clinical coding systems, drug formularies, and regulatory frameworks, not just translated versions of US-centric datasets.
Researchers at ETH Zurich working on multilingual clinical NLP have demonstrated that smaller models trained from scratch on linguistically appropriate data consistently outperform larger English-optimised models on local-language clinical tasks. This is not a surprising finding to anyone who takes architectural mismatch seriously, but it remains underweighted in procurement decisions by European health systems that default to large commercial models because of their English-language benchmark performance.
The European Health Data Space regulation, currently progressing through implementation, creates a framework for sharing health data across EU member states for secondary use including AI development. If the data shared under that framework remains predominantly in majority languages, the models trained on it will reproduce existing linguistic inequalities at scale. The technical standards being developed under EHDS implementation must include explicit requirements for linguistic representativeness, not just data volume and interoperability format.
Investment, Accountability, and the Path Forward
The gap between what European AI policy aspires to and what European AI deployment currently delivers for linguistic minorities is wide. The AI Act creates accountability mechanisms; they need to be used. National competent authorities designated under the Act must be equipped and willing to assess training data representativeness as part of conformity assessment for high-risk healthcare AI. Self-certification by developers, the default path for most high-risk AI systems, is insufficient where structural data bias is this entrenched.
Investment is equally necessary. The EU's Horizon Europe programme and the UK's AI Safety Institute both fund multilingual AI research, but funding for genuinely low-resource European languages remains modest relative to the scale of the problem. Public health systems procuring AI tools must begin demanding linguistic performance disaggregation as a standard component of procurement specifications: not just overall accuracy, but accuracy broken down by language, register, and demographic group.
The architecture of AI systems reflects the values and priorities embedded in their creation. Systems built on English-dominant data, with English morphological assumptions and Western institutional knowledge, will serve English speakers in Western institutional contexts most reliably. Serving European patients equitably, including the 60-plus million EU citizens whose first language is a regional or minority language, requires treating linguistic inclusivity as a first-order design requirement, not an afterthought applied through post-hoc fine-tuning.
Comments
Sign in to join the conversation. Be civil, be specific, link your sources.