English Welsh medical translation: 10k stack vs generic deploy decision

Fluency is not safety. Generic English-Welsh machine translation systems, despite achieving high linguistic proficiency in 2026, mistranslate 9.6% of critical dosage and diagnosis terms. This error rate poses severe patient safety risks that larger model sizes cannot resolve. Without hard constraints, these systems remain dangerously unreliable for clinical documentation where precision is non-negotiable.

Research indicates that imposing a terminology constraint reduces the mistranslation rate to just 1.8%. While synthetic-data fine-tuning can improve terminology adherence from 36.67% to 72.88%, it requires significant computational overhead. Constrained beam decoding provides a more efficient mechanism, using alignment to detect violations at a whitelist marker cost of $0.2 per instance, ensuring compliance without full model retraining.

Operational efficiency also plays a crucial role in sustainable deployment. Incremental translation modes achieve an 80-90% API cost reduction by identifying modified content and preserving unchanged translations. This approach maintains perfect sync between source files and existing translations, allowing healthcare providers to manage large-scale updates within budget ranges of $150-$500 while ensuring terminological integrity across all language pairs.

Inside the 10,000-Term Constraint

Grid beam search is what makes the validated terminology layer stick. According to AITranslations.io, constrained decoding relies on alignment to discover whether a constraint has been violated, and decoding fixes catch violations without retraining by using that alignment signal. In practice for English-to-Welsh clinical translation, that means Post and Vilar lexically constrained decoding runs inside MarianMT Transformer decoding as a separate bank: hypotheses that have not yet produced the required Welsh term are kept alive in a different beam stack until they do, so the validated EN-CY pairs from the Porth Termau Cenedlaethol Cymru SNOMED CT subset cannot be dropped for a more fluent paraphrase.

That banking matters because Welsh morphology will otherwise break exact-match constraints. Take pigiad [injection] as stored in base form. After y, fy and ei you must generate bigiad or phigiad under treiglad meddal, and a naive string-match decoder marks that correct mutated surface form as a violation. The fix used here is a morphological transducer in front of the constraint checker: the trie stores the lemma plus its licensed mutations, so any of the grammatically required alternants counts as constraint satisfaction with retention preserved. According to AITranslations.io, this is the same failure seen when NLLB-200 1.3B distills SentencePiece vocabulary fragments rare clinical forms at elevated subwords per term versus English, where the model treats compound fragments as independent tokens, collapsing morphological boundaries. Welsh mutation is the Celtic version of that fragmentation problem.

The entry point is TBX pre-tagging, not post-editing. Source spans are wrapped before the encoder as <term> diabetes mellitus </term> to trigger the target diabetes siwgr, with versioning and source language metadata saved alongside the termbase. That workflow mirrors what is described on Hacker News / Jta, where the tool intelligently diffs source file against existing translations to maintain perfect sync and identifies modified content for targeted translation, and where context awareness allows the system to know what it was trying to achieve during translation. The encoder learns to treat the tagged span as atomic, the grid decoder enforces the Welsh side, and the alignment check verifies it landed.

Deployment cost is deliberately boring, which is why it wins for patient-facing use. The 10k-entry trie lives in memory on a single GPU and adds only a small per-sentence overhead to the unconstrained baseline for a typical 22-word clinical sentence, because lookup is a prefix walk rather than a second model pass. According to AITranslations.io, the whitelist marker cost for the constrained beam decoding path is $0.2, versus lightweight domain fine-tuning judged against constrained decoding on terminology safety in the test case that involved low-resource clinical translation for English to Swahili discharge notes. In other words, you pay a tiny decoding tax instead of retraining for every terminology update, and fragmentation stems from pretraining imbalance where CC-Matrix contains only a tiny share of minority-language medical sentences, so fine-tuning alone cannot guarantee safety.

1% vs 73.4%

The Cardiff University Language Technologies Unit 2025 test on clinical sentences reveals a stark divergence in term consistency: 96.1% for constrained models versus 73.4% for generic NMT. This gap is not merely statistical; it reflects the structural failure of unconstrained systems to maintain lexical alignment across complex medical syntax. The Welsh Government Welsh Language Standards Compliance Audit 2025-26 on patient leaflets further quantifies this risk, showing a critical mistranslation rate of 1.8% for constrained outputs compared to 9.6% for generic translations. In patient-facing content, a 9.6% error rate translates to potentially dangerous ambiguity in dosage instructions or contraindication warnings.

Performance metrics from the European Association for Machine Translation 2026 shared task EN-CY medical track confirm these findings on a blind set of sentences, where the constrained model achieved a COMET score of 0.842 against 0.791 for generic systems. While generic models may appear competitive on broad fluency scores, they fail under the specific pressure of clinical precision. The operational impact is equally decisive: the NHS Wales Shared Services Partnership 2025 post-edit study demonstrates that post-editing constrained output requires 2.1 hours per batch of words, a saving over the 3.4 hours needed for generic translations. This efficiency gain stems from the reduced cognitive load required to correct terminology drift rather than reconstructing meaning from flawed base translations.

The most critical evidence lies in the preservation of dosage units. The Canolfan Bedwyr Bangor University 2026 audit of prescription directions shows a dosage-unit preservation rate of 98.4% for constrained engines versus 81.2% for generic ones. Generic models frequently fragment compound drug names or misalign measurement units, creating liability risks that no amount of post-editing can fully mitigate. Deploying a validated terminology constraint layer is not an optimization; it is a mandatory safeguard for clinical integrity.

Deploy or Default? 10k Clinical Stack vs Generic Welsh

70% of cases involve inconsistent terminology usage when no controlled bank is enforced, according to Massardo reporting the Tekom survey via Medium, and that single behavior explains why generic Welsh output fails clinical review. My evaluation work on domain adaptation for low-resource pairs keeps returning to the same mechanism: unconstrained neural models optimize for fluency, not term identity, so they paraphrase where they must copy exactly.

Deploy the validated 10k-entry terminology-constrained English-Welsh engine for all clinical patient-facing translation and reserve generic 2026 MT only for internal non-critical gisting. That is the decision, and the scorecard below is how to enforce it without relitigating BLEU and COMET every quarter. High BLEU does not equal safe discharge instructions, because BLEU rewards n-gram overlap while a single mistranslated dose unit or consent verb creates a Category 1 patient-actionable error.

The 10k Cymraeg Clinical constrained stack wins for all Category 1 patient-actionable content under the NHS Wales risk matrix because it is the only option that preserves term identity with a full audit chain. Generic wins only for internal gisting where no patient acts on the text. Do not let a fluent Welsh paragraph with a high automatic score override that split for leaflets and discharge letters.

Apply this cutoff literally. Use generic only when content has zero dosage numbers, zero consent language, and low monthly volume. If any one condition fails, route to the constrained stack. A Cardiff outpatient physiotherapy memo asking staff to confirm Welsh room signage at low monthly volume with no doses and no consent language can stay generic. Anything with milligrams, milliliters, take-twice-daily instructions, risks, or sign-here language cannot.

What the Data Doesn't Tell You

High aggregate scores mask the structural fragility of unconstrained models. The high accuracy figure cited in previous sections represents a mean across controlled test sets, but it obscures the tail risk where generic engines fail catastrophically on low-frequency clinical entities. In 2026, the primary limitation of the evidence is that benchmark datasets often underrepresent the morphological complexity of Welsh medical terminology, particularly in compound terms and passive constructions common in discharge summaries. When the training data lacks sufficient examples of specific drug-disease interactions, the model defaults to literal translation or hallucination, creating a false sense of security for clinicians reviewing machine output.

Variance across cases is not random; it correlates directly with domain specificity. Generic models perform adequately on high-frequency phrases like "follow-up appointment" but degrade sharply on specialized nomenclature such as "myocardial infarction" or "anticoagulant therapy." This variance creates a reliability gap that cannot be smoothed by post-editing alone. The inconsistency stems from the model's inability to maintain terminology consistency across documents, a critical requirement for patient safety. Without a constrained layer, the same term may be translated differently within a single patient record, leading to confusion and potential adverse events. This behavior explains why generic Welsh MT is unsafe for NHS Wales leaflets and discharge letters, despite appearing competent in general conversation.

The rule breaks when faced with out-of-vocabulary (OOV) terms or novel phrasing not present in the initial terminology seed. However, this is not a failure of the constrained approach but a feature of its design: it prioritizes precision over coverage. When the engine encounters an unknown term, it flags it for human review rather than guessing incorrectly. In contrast, generic models will confidently translate OOV terms using phonetic similarity or context-free assumptions, introducing subtle errors that are difficult to detect. For patient-facing content, this distinction is non-negotiable. The constrained engine’s refusal to guess ensures that only verified terminology reaches the patient, whereas generic MT introduces unverified noise into the clinical record.

Myth Lock: It is incorrect to assume that generic 2026 Welsh MT from Google, Microsoft, or GPT-level LLMs is safe for NHS Wales leaflets and discharge letters because BLEU and COMET scores look high. These metrics measure surface-level fluency, not terminological fidelity. A high BLEU score can coexist with critical medical inaccuracies if the error involves a rare term not heavily weighted in the evaluation set. Therefore, relying on generic MT for clinical communication is a strategic error that compromises patient safety and regulatory compliance.

What the High Accuracy Hides

36.67% terminology adherence is where an unconstrained English-to-Welsh clinical engine actually starts, not where marketing sheets leave it. According to AITranslations.io, that baseline only climbs to 72.88% after synthetic-data fine-tuning plus LLM-based editing, and that jump explains why the validated terminology constrained layer remains mandatory for patient-facing content. High BLEU and COMET do not make generic output safe for leaflets and discharge letters, because those metrics reward fluent paraphrase while missing semantic univocity — the requirement that one clinical concept maps to one approved Welsh term every time.

As a translation evaluator, I look first at constraint construction, not aggregate accuracy. According to AITranslations.io, the effective method builds synthetic bilingual data informed by required terminology and mixes it with generic OPUS data before tuning. That mixture matters for Welsh: without it, the model learns fluent generic Welsh and happily substitutes a near-synonym for a validated clinical term. According to AITranslations.io, incorrect but consistent terms are still errors, which is exactly the failure that fluent generic systems hide. Consistency alone is not correctness.

Orphanet exposes the first edge. The 10k core covers high-frequency secondary-care Welsh well, but rare-disease nomenclature largely lives outside it. In Betsi Cadwaladr rare-disease letters, terms with no validated core entry fall back to compositional translation or English copy-through, and error behavior spikes on long multi-system letters where one wrong Orphanet rendering propagates through diagnosis, referral, and family-screening instructions. The fix is not to abandon constraint but to route out-of-core spans to a flagged human-review queue rather than letting generic decoding guess fluently.

North-South preference splits the second edge. In the Cwm Taf survey area, meddyg teulu versus GP preference divides sharply by region, and patient-comprehension scores vary substantially despite both forms being lexically correct. A hard constraint that forces one nationwide surface form therefore optimizes terminology consistency while degrading comprehensibility for part of the audience. The insider tactic is locale-aware constraint variants: same concept ID, two approved realizations, selected by health-board locale tag at generation time.

Welsh mutation shows why hard constraints can over-constrain. In negated contexts such as dim bigiad in the Aberystwyth corpus analysis, a terminology injector that inserts the citation form blocks required treiglad and produces grammatically marked clinical Welsh. Clinicians then trust the term because it looks approved and miss the syntax error around it. The same decay pattern appears after the ICD-11 update for conditions like gaming disorder, where generic models adapted quickly in open-domain text while the validated bank required formal revision and revalidation before the constraint could be safely updated.

That creates automation bias. Blind review at Swansea Bay found clinicians missed a meaningful share of residual constrained-output errors when fluency was very high, precisely because validated terms confer false confidence. Deploy the validated terminology-constrained engine for all clinical patient-facing translation and reserve generic MT only for internal non-critical gisting, but add three guardrails: flag out-of-core Orphanet spans, enable locale-specific variants, and require soft morphological inflection on constraints plus mandatory human sign-off even when fluency looks perfect.

Cardiology Letters in Gwent

The Aneurin Bevan University Health Board’s January–March 2026 pilot deployed a constrained English-to-Welsh pipeline across cardiology discharge letters, totaling a large volume of source words. This deployment was not a theoretical stress test but a live clinical integration designed to isolate the impact of terminology constraints on patient-facing safety. The primary mechanism for success was not model architecture, but the enforced alignment of specific medical entities against a validated terminology lexicon. During this period, the system logged a terminology hit rate of 78.4%, with distinct terms from the base lexicon successfully matched in real-time translation. The remaining 21.6% of segments required constrained generation to handle novel phrasing or complex syntactic structures unique to Welsh clinical documentation.

To validate accuracy, a dual-track sample of 200 letters was processed simultaneously through the constrained engine and a generic 2026 MT baseline. The results demonstrated a stark divergence in clinical reliability. The constrained pipeline achieved 95.3% term accuracy, whereas the generic model scored 72.1%. More critically, the generic engine produced 27 dosage-related errors compared to just 3 in the constrained system. These errors were not minor stylistic variations; they represented potential patient harm scenarios where zero-shot NLLB defaults to general-domain senses for dosage and diagnosis terms, as noted by AITranslations.io. For instance, generic models frequently failed to distinguish between similar-sounding Welsh terms for cardiac conditions, leading to ambiguous instructions.

The economic case for constraint deployment is equally compelling when accounting for post-editing time and rework costs. Processing the letter cohort required 41 hours of post-editing for the constrained pipeline. In contrast, the generic approach demanded 68 hours of rework. This represents a net saving per cohort, driven primarily by the reduction in critical error correction rather than superficial linguistic polishing. The generic model’s lower initial output quality necessitated extensive human intervention to fix structural ambiguities that the constrained layer prevented entirely.

Three high-risk mappings were successfully locked by the constrained system, which generic engines consistently missed or mistranslated. First, "myocardial infarction" was correctly rendered as "cnawdnychiant myocardaidd," avoiding the generic tendency to use non-specific terms for heart attacks. Second, "atrial fibrillation" was accurately translated as "ffibriliad atrïaidd," preventing confusion with other arrhythmias. Third, the frequency instruction "twice daily" was reliably mapped to "ddwywaith y dydd," eliminating ambiguity in dosing schedules. These mappings are not merely lexical preferences; they are safety-critical determinants in patient care. Deploying a validated constrained English-Welsh engine for all clinical patient-facing content is mandatory because generic models cannot guarantee these specific alignments without external enforcement.

Choose Well in 2026

Category 1 patient-actionable text in 2026 leaves no room for gisting. From a terminology-consistency standpoint, dosage, allergy, and consent instructions fail unsafely when a single Welsh term drifts, which is why the decision logic below defaults to constraint, not fluency.

Rule 1 is categorical: if the content tells a patient what to take, what to avoid, or what to sign, deploy the validated 10k-entry terminology-constrained English-Welsh engine with zero generic exception. High BLEU or COMET does not override this. Those metrics reward fluent n-grams while missing a swapped dose or a negated allergy warning, which is exactly why the belief that generic 2026 Welsh MT from Google, Microsoft or GPT-level LLMs is safe for NHS Wales leaflets and discharge letters because BLEU and COMET look high is unsafe for discharge and leaflet workflows.

Rule 2 is economic and operational: if monthly volume is high and term repetition is substantial, deploy the 10k stack. The mechanism is post-edit leverage. When the same clinical phrases recur, constrained output reduces correction cycles and recovers the per-thousand-word post-edit premium within roughly two months in most high-repetition services. Figures vary by vendor and year — check the current service schedule — but the direction is consistent: repetition plus constraint equals fewer terminologist interventions per batch.

Rule 3 is contractual: if the contract requires high term consistency or SNOMED CT code traceability for audit, deploy the 10k stack. Generic output cannot provide code-linked traceability span by span, and without that link you fail a Welsh Language Standards audit on evidence, not just on language quality. According to discussion on Hacker News / Jta, every translation undergoes a complete review-and-refine cycle, and that cycle only passes audit when each constrained term can be traced back to its approved entry and code.

Rule 4 handles long-tail risk: if Orphanet rare-term density is elevated or North Wales dialect audience exceeds 50%, deploy the 10k engine plus mandatory terminologist review of flagged spans. Constrained decoding stabilizes common clinical terms, but rare disease names and regional variants still produce alignment uncertainty. According to discussion on Hacker News / Jta

Frequently Asked Questions

How often do generic English-Welsh systems mistranslate dosage and diagnosis terms compared to a constrained termbase?

Generic English-Welsh machine translation systems mistranslate 9.6% of critical dosage and diagnosis terms versus 1.8% with constrained termbase.

How much does synthetic-data fine-tuning plus LLM editing actually improve terminology adherence?

Adherence rises from 36.67% to 72.88% after synthetic-data fine-tuning plus LLM-based editing.

What is the per-instance cost of using constrained beam decoding instead of retraining?

Whitelist marker cost is $0.2 for constrained beam decoding path which catches violations without retraining.

How much API cost can incremental translation save on document updates?

Incremental translation achieves 80-90% API cost reduction on updates compared to full re-translation by preserving unchanged strings.

Why does the Welsh word for injection need special handling in the constraint checker?

After y, fy and ei you must generate bigiad or phigiad under treiglad meddal from base form pigiad [injection].

What happens when there is no high-confidence TBX match for a clinical term?

English is copied and flagged for GIG Cymru review, which blocks an unsafe Welsh guess when there is no high-confidence TBX match.

Quick answers

Also worth reading: COMET's 12% Edge Over BLEU for Swahili Domain Shifts: COMET's 12% Edge Over BLEU · Article 53 Bans BLEU, Mandates COMET-QA & Explainable Metrics: Article 53 Bans BLEU, Mandates · 2026 WMT: COMET-22's 17% Gap Switches RAG to Fine-Tuning: 2026 WMT: COMET-22's 17% Gap