A glossary is not enough to guarantee terminology

The term is present. The sentence is wrong.

You have an approved glossary. You give it to the translation engine. You check the output for the required words. Every word is there. That can still be a failed translation.

A glossary usually gives you the form you would look up in a dictionary. A sentence needs the form that fits its grammar. Xu and Carpuat illustrate the difference with the Lithuanian word for violin: the glossary supplies smuikas, while buying the violin requires smuiką in their example. The change is grammatical, not a departure from the approved terminology.¹

Glossary:  violin -> smuikasSource:    They bought the violinTarget:    Jie nusipirko smuiką

An exact search for the glossary form rejects that target. Replacing the inflected word with the dictionary form would make the search pass by damaging the sentence. The check and the requirement have come apart.

We have written about retrieving the right translation memory match. A glossary introduces a different obligation: use the approved term even when a retrieved example suggests another. But before that obligation becomes a gate, you have to define what counts as using it.

The glossary entry has to apply first

Consider an illustrative English interface with a glossary entry for account. The product uses that word for a user account. Now the same catalog contains the sentence “Take delivery time into account.” A literal source match finds the entry. The product concept is absent.

Forcing the approved user-account translation into that sentence enforces the wrong decision. Better target-side inflection cannot repair a source-side sense error. The pipeline needs to establish that the entry applies before making its target mandatory.

Record the intended concept, the domain where it applies, and an example of its use. Part of speech matters where the same spelling serves different roles. A termbase can carry this information, but exporting only its source and target columns discards it. If the remaining text cannot settle the sense, preserve that uncertainty for review instead of silently turning a candidate into a requirement.

Overlapping entries need a policy too. In an illustrative software glossary, key and API key may have different approved translations. Matching both and enforcing both independently can ask the target sentence to express the same source span twice. Select the applicable phrase before constructing its constraint.

A hard constraint guarantees the condition you encoded

Presence and correct use are separate measurements. Bane and colleagues evaluated terminology methods by asking linguists about term presence, grammatical form, agreement in surrounding words, duplication, accuracy, and fluency. In their English-to-Russian evaluation, alignment-based insertion was strongest at including the term and weakest at morphological agreement. The Japanese-to-English results differed. Their small human evaluation reported raw results without significance testing.²

That is a concrete reason to distrust a single glossary-adherence score. A system can improve the presence measurement while making the surrounding translation worse.

Constrained decoding changes the search itself. Hasler and colleagues use finite-state machines and multiple decoding stacks to track lexical requirements. They also connect constraints to source spans through attention, addressing misplaced terms and duplicated translations. Supplying target words alone leaves their relationship to the source insufficiently specified.³

For a feasible set of exact lexical constraints, a decoder can require a completed candidate to contain those token sequences. It has then enforced token inclusion. It has not proved that the selected term expressed the intended concept or agreed with its neighbours. If search cannot produce a valid candidate within its limits, that must remain a failure, not become permission to return unchecked output.

The dictionary form is an input to the decision

One response is to teach the model how to incorporate terminology. Dinu and colleagues trained NMT with inline target-term annotations and source factors that distinguish those annotations from running text. Their approach learns to use the supplied terms without adding constrained-search overhead at inference. That is learned behaviour, so the output still needs checking.⁴

Another response makes inflection an explicit step. Xu and Carpuat predict a target term's grammatical form from source context, then pass that form to translation. Their framework separates morphological decisions from the translation model and evaluates rule-based and neural inflection modules on English-to-German health and English-to-Lithuanian news text.¹

The useful engineering distinction is between fixing a lexical choice and freezing its spelling. A policy may require a particular term while allowing the inflections that make it grammatical. An identifier may instead require exact preservation. Those are different contracts and should not share an undifferentiated match setting.

Lemmatization can help recognize an allowed form. It cannot, by itself, prove that the form is correct in this sentence. Recognizing that a word belongs to the approved lemma leaves case, number, and agreement to be checked. Accepting every form of the right word would merely move the false assurance from string matching to morphology.

A better prompt still needs an acceptance check

An LLM can receive the glossary in a prompt or revise a translation after generation. Both are useful ways to produce a candidate. Neither makes the instruction self-enforcing.

A 2026 preprint by Sánchez-Torrón and colleagues compares expert prompts with automatically optimized prompts for tasks including glossary insertion into existing translations. Optimized prompts improved term match rates for several models without consistently improving fluency. Their term match rate counts stems shared with glossary terms in the gold translation; it is not a measure of full grammatical correctness. The experiments use proprietary datasets.⁵

A correction pass also creates a fresh translation to validate. Suppose a repair replaces a noun with the approved synonym. Its article and adjective may now need changing. If the editor rewrites the clause to accommodate them, it may affect another term or a protected placeholder. Checking only the originally missing word leaves those edits unexamined.

Treat the repaired sentence as a new candidate. Run the applicable checks again, including the structural checks that were already satisfied before editing. Give retries a limit and route unresolved cases to review. Another model call is an attempt to meet the conditions, not evidence that they were met.

Enforce a policy you can actually check

The following is an engineering recommendation drawn from those distinctions. Keep the acceptance conditions explicit, and report the evidence for each separately.

Start with the rules that admit a precise test: exact identifiers, explicitly forbidden variants within a known scope, and protected message structure. For linguistic checks, record whether the result is established, violated, or unresolved. A checker that cannot decide must not quietly return the same status as a checker that found the required evidence.

Keep the glossary revision and the applied entry with the decision. Otherwise, when an approved term changes, you cannot distinguish an old acceptance under an earlier policy from a failure to apply the current one. A retrieved, previously approved translation needs the same current-policy check as a newly generated translation.

Measure missed terms separately from wrongly applied terms and grammatical defects. A system that inserts every matched glossary entry can look excellent on omissions while enforcing irrelevant meanings. The review set needs examples where the right action is to leave an entry unused, as well as examples where the correct target has a different surface form from the glossary.

A glossary records the lexical decision. The pipeline must determine where it applies, realize it grammatically, and enforce the release conditions on the final sentence. Uploading the glossary supplies the beginning of that process.

References

1. 1.Weijia Xu and Marine Carpuat, 2021 Rule-based Morphological Inflection Improves Neural Terminology Translation EMNLP 2021
2. 2.Fred Bane, Anna Zaretskaya, Tània Blanch Miró, Celia Soler Uguet and João Torres, 2023 Coming to Terms with Glossary Enforcement: A Study of Three Approaches to Enforcing Terminology in NMT EAMT 2023
3. 3.Eva Hasler, Adrià de Gispert, Gonzalo Iglesias and Bill Byrne, 2018 Neural Machine Translation Decoding with Terminology Constraints arXiv:1805.03750
4. 4.Georgiana Dinu, Prashant Mathur, Marcello Federico and Yaser Al-Onaizan, 2019 Training Neural Machine Translation To Apply Terminology Constraints arXiv:1906.01105
5. 5.Marina Sánchez-Torrón, Daria Akselrod and Jason Rauchwerk, 2026 To Write or to Automate Linguistic Prompts, That Is the Question arXiv:2603.25169v2, preprint