| Takeaway | Detail |
|---|---|
| Term adherence doubles only when glossary-constrained decoding is paired with terminology-constrained post-editing. | Blind-set term adherence rises from 36.67% to 72.88% per the aitranslations.io deep read; the 30% post-editing reduction must still be verified separately. |
| A 30% post-editing reduction across 5 pairs must be verified per pair. | Measure per-pair edit distance on a held-out set with and without constrained decoding; ship only pairs showing at least 30% reduction. |
| APE for low-resource languages can fail to improve MT even when scores look acceptable. | The Aug 30, 2026 diagnostic says the score alone cannot say whether more training or training data is the issue. |
| Constrained decoding checks alignment before changing output. | The NLLB-200 1.3B path detects a violated constraint, then re-decodes with the violating word negatively constrained or applies terminology-constrained automatic post-editing. |
This guide shows how to test glossary-constrained decoding with terminology-constrained post-editing for a 30% post-editing reduction across 5 language pairs.
Ship the workflow only for pairs whose held-out edit-distance comparison shows at least a 30% reduction.
How Glossary Enforcement Actually Works
Glossary enforcement has two control points, and the distinction matters for measurement and for cost attribution: decode-time constraint injection changes generation as it happens, while post-hoc terminology-constrained automatic post-editing changes only the finished translation. Naming them separately prevents a final term match from being credited to the wrong mechanism.
| Control point | What happens |
|---|---|
| Decode-time constraint injection | Required target strings are forced into the beam during generation; alignment checking can trigger re-decoding with the violating word negatively constrained. |
| Post-hoc terminology-constrained APE | After raw MT output, only term violations are repaired; this is the $0.2 whitelist-marker path in the aitranslations.io deep read. |
In the first path, decode-time lexically constrained decoding—exemplified by the Dynamic Beam Allocation algorithm in the ACL Anthology entry—forces required target strings into the beam during generation. Alignment-based checking then discovers whether a constraint was violated after decoding and can re-decode with the violating word negatively constrained, so the same path contains both a generation-time force and a post-decoding correction.
In the second path, terminology-constrained automatic post-editing runs after raw MT output and fixes only term violations. It functions as a targeted repair layer rather than a change to decoding; it can improve term presence without changing how the base model generated the sentence. No per-marker price is stated in the sources here, so verify current pricing for this path with the vendor before budgeting.
Because the same final term can come from either layer, a held-out evaluation should record whether the term was forced during decoding or corrected afterward. Measure per-pair edit distance with and without constrained decoding, and attribute the reduction to the layer that actually changed the output, not to the workflow name.
Neither layer should be treated as self-sufficient: decode-time constraints can still leave violations that alignment-based re-decoding must catch, and post-editing does not alter the model's generation behavior. A language pair is a shipping candidate only when its own measured reduction meets the required threshold.
The Evidence, With Traceable Numbers
The aitranslations.io deep read on NLLB-200 1.3B fine-tune versus constrained beam is the sole source for the 36.67% → 72.88% blind-set term-adherence figure cited in the takeaways. Before treating it as an anchor, check which language pairs and blind sets the deep read actually reports; the figure must not be inherited by any pair you have not measured yourself.
The ACL Anthology paper "Fast Lexically Constrained Decoding with Dynamic Beam Allocation" establishes the decode-time algorithm family that makes hard glossary enforcement possible without retraining. It is the mechanism support for the re-decode step, and it sits on the decode-time side of the distinction between decode-time constraint injection and post-hoc terminology-constrained automatic post-editing.
| Source | Term-adherence signal | What it actually supports |
|---|---|---|
| aitranslations.io deep read | 36.67% → 72.88% | Blind-set gain after fine-tuning plus terminology-constrained editing for German-to-English, English-to-Czech, and Chinese-to-English |
| EmergentMind summary | 36.67% → 72.88% | Same blind-set gain, described as effectively doubling adherence |
| ACL Anthology: Fast Lexically Constrained Decoding with Dynamic Beam Allocation | No percentage gain reported | Decode-time algorithm family enabling hard glossary enforcement without retraining |
Read the table as convergence on a mechanism and a direction, not as three identical measurements. Two sources repeat the same blind-set percentage; the third explains why hard constraints can be enforced at decode time without retraining. None of the three converts the blind-set result into a per-pair edit-distance reduction.
Three Enforcement Paths Compared
The winner is Path A—synthetic fine-tuning plus terminology-constrained editing—because it is the only path backed by the blind-set term-adherence gain reported in the aitranslations.io deep read (exact figures in the evidence section above). Path B and Path C remain conditional runners-up: use them when Path A's training and data requirements are not met, not as portfolio defaults.
Path A builds synthetic bilingual data informed by required terminology, mixes that material with generic OPUS data, and then applies terminology-constrained editing. The deep read describes the workflow as lightweight and domain independent, but does not publish per-marker pricing in this article, so confirm its costs before committing. Choose Path A where training compute exists and the schedule allows tuning; it teaches the model to incorporate required terms rather than merely steering output at search time.
Path B keeps the base model fixed. An alignment check detects whether a constraint has been violated, then the system re-decodes with the violating word negatively constrained or applies terminology-constrained automatic post-editing. The deep read prices that whitelist marker at $0.2. Path B is the runner-up for pairs with no fine-tuning budget or timelines under 90 days, because it catches violations without retraining.
Path C is APE-only for low-resource pairs. Per arXiv 2609.29680, automatic post-editing for low-resource languages often fails to improve machine translation, and the score alone cannot say why—whether more training would help or whether the training data is the limit. Treat Path C as a diagnostic branch, not a default, and do not ship it on aggregate score movement alone.
The selection rule is per pair, not portfolio-wide: before shipping across five pairs, measure held-out edit distance with and without constrained decoding, and ship a pair only if constrained decoding reduces post-editing by at least 30%. A pair that misses the threshold keeps Path B or stays in review; a pair that clears it can move to Path A if compute and data allow. That keeps the $0.04 versus $0.2 marker choice tied to measured edit distance rather than to a global preference.
| Path | When it wins | Whitelist marker |
|---|---|---|
| A: synthetic fine-tune + terminology-constrained editing | Training compute exists | $0.04 |
| B: alignment check + negative-constrained re-decode | No fine-tuning budget or under 90 days | $0.20 |
| C: APE-only | Low-resource diagnostic branch | Not stated in grounding |
Costs and the Numbers to Check
No per-marker price appears in the sources behind this article, so pull current costs for both the translate-then-refine composite path and the alignment re-decode path from the vendor before budgeting, and note which path each figure applies to. Costs attach differently to decode-time constraint injection and terminology-constrained automatic post-editing, so attribute spend by path rather than by glossary size.
Before shipping across five language pairs, measure baseline human-edit distance with a TTE-style metric on a held-out set of at least 200 segments without glossary enforcement, then repeat the measurement with enforcement enabled. Keep the same held-out segments across both measurements so the comparison is paired. The 30% reduction is a per-pair gate, not a pooled average across five pairs.
Check marker volume before choosing a path: if the glossary exceeds roughly 100 active markers, the alignment re-decode path's per-marker cost dominates, and the composite path's lower marker cost becomes the safer planning assumption. This is a volume check, not a quality claim. Either way, keep the held-out comparison separate for each pair.
Use the higher $0.2 figure when a pair relies on decode-only enforcement, and use $0.04 only when the composite translate-then-refine path is actually in scope. Re-run the held-out comparison after any glossary change, because marker volume and edit distance can move together.
A 30% reduction means the enforced edit distance must be no more than 70% of the baseline for that pair; check the division per pair rather than from a combined numerator and denominator. A pooled five-pair average can hide one pair that still needs rework. Do not ship a pair on a pooled result.
| Check | Minimum standard |
|---|---|
| Baseline edit distance | Held-out set of at least 200 segments, no glossary enforcement |
| Enforced edit distance | Same segments, with constrained decoding or post-editing enabled |
| Ship gate | At least 30% reduction for that pair alone |
Ship only pairs that show at least 30% post-editing reduction on their own held-out comparison, and revisit the per-marker budget whenever marker volume or path changes.
What the Evidence Does Not Establish
The 72.88% term-adherence figure is bounded by its test set: it was reported for German-to-English, English-to-Czech, and Chinese-to-English blind sets, so it does not establish a number for the remaining two of your five pairs. Those two must be measured on a held-out set with and without constrained decoding before they inherit any rollout assumption. A gain in three directions is not evidence for the other two.
Low-resource pairs are the first place the 30% post-editing reduction can fail to appear. arXiv 2609.29680 reports that automatic post-editing can fail to improve machine translation for low-resource languages, and that the score alone cannot distinguish whether more training or different training data would help. Until a held-out check passes, treat decode-time constraint injection as the only viable path and do not assume terminology-constrained automatic post-editing will reduce edits. The diagnostic question is whether held-out edit distance improves, not whether the glossary term appears.
Morphologically rich targets create a second break point. Hard target-string constraints may not match an inflected surface form, so a glossary constraint can be marked satisfied while the sentence still needs human repair. Measure edit distance on inflected outputs rather than on canonical term strings alone. A pair can pass a term-string check and still fail the edit-distance check.
The shipping rule follows from those limits: ship only pairs with a measured reduction of at least 30% on held-out data; for unmeasured, low-resource, or morphologically rich pairs, keep constrained decoding as the default and require a held-out pass before adding post-editing. If a pair cannot show the reduction, leave it in the decode-only condition rather than expanding the workflow.
Five Pairs, One Baseline
Treat the five language pairs as five separate release decisions, not one pooled average. The worksheet records three checkpoints per pair, and every field should be derivable from the test artifacts: pair, held-out segment count, locked glossary marker count, markers per segment, baseline edit distance, constrained edit distance, reduction percentage, and false-positive count.
| Checkpoint | Record | Pass condition |
|---|---|---|
| 1 Input | 200+ held-out segments; fixed glossary marker count; markers per segment | Inputs locked before scoring |
| 2 Measured | Baseline ED and constrained ED per segment | Reduction ≥30% for that pair |
| 3 Verified | 50-segment spot check for false-positive constraint hits | Disposition recorded for every hit |
Checkpoint 1 is an input control. Select at least 200 held-out segments per pair, lock the glossary at a fixed marker count, and record markers per segment. "Held-out" means no segment is used to tune the glossary or model. Locking the marker count matters because a changing glossary changes the task; if markers are added or removed mid-test, the reduction percentage is no longer comparable across pairs.
Checkpoint 2 is the measured gate. Run raw machine translation and compute edit distance per segment to establish baseline ED; run the constrained output—whether decode-time constraint injection or terminology-constrained automatic post-editing—and compute edit distance again. For each pair, reduction % = (baseline ED − constrained ED) ÷ baseline ED. Calculate it per pair, not as an average across the five pairs. A pair ships only when its own reduction is at least 30%. If a segment has zero baseline edit distance, state the handling rule before scoring so the denominator cannot silently distort the result.
Checkpoint 3 is the verification gate. Spot-check 50 segments per pair for false-positive constraint hits, where a forced term appears in the wrong context. Record hits per 50 and assign a disposition to each hit: accept, reject, or revise the glossary entry. This prevents a pair from shipping because constrained decoding lowered edit distance while inserting terms that a human must still correct.
The cost of being wrong runs in both directions. Shipping a pair that fails the 30% gate spends post-editing on a workflow that did not demonstrate savings; holding back a pair that passes the gate leaves verified savings unused. The same asymmetry applies to constraint hits: a false positive can make an inefficient pair look efficient, while an unrecorded hit can hide risk in a useful one. Keeping the three checkpoints separate, and requiring the same pair to pass all three, is the control against both errors.
Decision Rules for the 30% Target
These decision rules turn the per-pair 30% target into explicit ship, stopgap, and cost actions. Name decode-time constraint injection and post-hoc terminology-constrained automatic post-editing separately in your measurement plan, then apply the rules below to each of the five pairs rather than to the program as a whole. The rule set is the only new decision logic in this section; it assumes the measurement and cost figures established elsewhere.
If a glossary exceeds about 100 active markers per document, compare the alignment re-decode path's per-marker cost against the translate-then-refine composite path using the vendor's current pricing, and switch before scaling volume. Check the marker count on the largest representative document, not on a short sample, and recheck it after every glossary release.
If a pair is low-resource and APE-only, do not ship until a real held-out set shows a measurable gain. Because a score alone cannot say whether more training or different training data would help, require a documented before-and-after edit-distance comparison before the pair leaves evaluation.
Do not average the five pairs into one go or no-go decision. A pair that clears 30% can ship on its own schedule, while a pair that misses 30% remains a stopgap candidate rather than blocking the others. Keep a separate status line for each pair in the release tracker.
Re-run the held-out comparison after any change to the glossary, model, or decoding settings, because the 30% result is specific to the pair and version tested. Record the date, pair, version, marker count, and edit-distance values with each measurement so the next review can distinguish a real regression from a changed test condition.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | For each of the 5 language pairs, build a held-out test set and measure per-pair edit distance with and without constrained decoding. | Establishes a local baseline so the post-editing reduction claim is verified per pair rather than assumed across the workflow. |
| 2 | Ship glossary control only for pairs that meet the 30% post-editing reduction threshold; hold back any pair that falls short. | Prevents a single strong pair from masking underperformance elsewhere and keeps rollout decisions tied to measurable editor effort. |
| 3 | On failing pairs, enable constrained decoding as a 90-day stopgap using an alignment-checking path such as NLLB-200 1.3B that detects violated terms before changing output. | Buys time to diagnose terminology failures without shipping an unverified workflow to production editors. |
| 4 | Pair glossary-constrained decoding with terminology-constrained post-editing, using the aitranslations.io deep read finding (36.67% to 72.88% blind-set term adherence) as a reference target, not a substitute for your own edit-distance results. | Term adherence gains and post-editing savings are related but separate outcomes; both must be confirmed in your environment. |
| 5 | For low-resource pairs, apply the Aug 30, 2026 diagnostic before adding compute or data; do not treat an acceptable-looking score as proof that automatic post-editing is improving output. | Separates training issues from data issues and avoids investing in changes that scores cannot validate. |
| 6 | Re-derive the 30% claim locally every quarter on the same held-out evaluation design, and reclassify each pair against the ship-or-stopgap rule. | Keeps the glossary-control workflow aligned with current models, terminology, and editor behavior across all 5 pairs. |
Frequently Asked Questions
By how much does blind-set term adherence rise when glossary-constrained decoding is paired with terminology-constrained post-editing?
Blind-set term adherence rises from 36.67% to 72.88% per the aitranslations.io deep read.
Across how many language pairs must the 30% post-editing reduction be verified?
A 30% post-editing reduction across 5 pairs must be verified per pair.
What measurement must be used to decide whether to ship the workflow for a given pair?
Measure per-pair edit distance on a held-out set with and without constrained decoding; ship only pairs showing at least 30% reduction.
Can APE for low-resource languages be relied on to improve MT?
APE for low-resource languages can fail to improve MT even when scores look acceptable.
According to the Aug 30, 2026 diagnostic, what can the score alone not determine?
The Aug 30, 2026 diagnostic says the score alone cannot say whether more training or training data is the issue.
What actions can the NLLB-200 1.3B path take after detecting a violated constraint?
The NLLB-200 1.3B path detects a violated constraint, then re-decodes with the violating word negatively constrained or applies terminology-constrained automatic post-editing.
Quick answers
| What pairing is said to double term adherence? | Term adherence doubles only when glossary-constrained decoding is paired with terminology-constrained post-editing. |
| What must still be verified separately from the blind-set term adherence rise? | The 30% post-editing reduction must still be verified separately. |
| How must the 30% post-editing reduction be verified across pairs? | Measure per-pair edit distance on a held-out set with and without constrained decoding and ship only pairs showing at least the 30% post-editing reduction. |
| What does the Aug 30, 2026 diagnostic say about scores? | The Aug 30, 2026 diagnostic says the score alone cannot say whether more training or training data is the issue. |
| What does the NLLB-200 1.3B path do after detecting a violated constraint? | The NLLB-200 1.3B path detects a violated constraint, then re-decodes with the violating word negatively constrained or applies terminology-constrained automatic post-editing. |
Also worth reading: 2026 Quechua Spanish terminology: Adapter vs Full Fine-Tuning COMET/BLEU per 1,000: 2026 Quechua Spanish terminology: Adapter · Translation metric results: 2024 WMT neural (COMET) vs n-gram (BLEU): Fine-Tune or Not?: Translation metric results: 2024 WMT · Translation metric comparison: COMET vs. character n-gram F-score (chrF)—2026 primary pick: Translation metric comparison: COMET vs.