When a medical authority renames a disease, retires a drug name, or revises an official technical term in Hindi or Tamil, the machine translation systems in use do not notice. They keep producing the old term, because it is what they saw thousands of times during training. Translation research has a well-developed toolkit for injecting required terminology at translation time. All of it assumes the glossary is a fixed list of correct pairs. It also tends to be tested on terms the model either never saw or already translates correctly. Separately, work on knowledge obsolescence in language models has found that models anchor on familiar, outdated facts even when the current fact sits right in front of them. Its sharpest result is that this anchoring persists under perfect retrieval (Can Agent Memory Systems Track Evolving State?).
This note proposes to bring those two findings together. The hard case for terminology translation is not the brand-new word. It is the superseded word: a term whose source side may be unchanged but whose correct target equivalent has been replaced. Here the model's strong prior actively works against the glossary. We will build a test set of real and controlled term supersessions and measure how often existing terminology methods leak the retired term. We also propose a simple training change: teach the model on constraints that deliberately contradict its prior, and mark the retired form explicitly so it can be suppressed. The goal is to close that leak.
Core Claim
Existing terminology methods report high term-usage rates. Those rates overstate how well the methods handle obsolete terminology, for two reasons. First, the constraints sampled during their training and testing are mostly ones the model already agrees with. In Training Neural Machine Translation to Apply Terminology Constraints, the unconstrained baseline already uses 76% of the required terms. Second, they measure whether the new term appears, not whether the old one disappears.
We predict that the failure rate on supersessions grows with how often the retired target term appeared in training. This is the translation analogue of the "salience trap" documented for agent memory. We also predict that two ingredients will flatten that curve without the cost of retraining for every update: training on prior-contradicting constraints, and a separate input marker for the retired form. If plain glossary injection already shows no frequency dependence, the claim is wrong. The experiments are built to show that cleanly.
Where This Sits
Three lines of work supply the machinery we reuse, and all treat the glossary as static. Training Neural Machine Translation to Apply Terminology Constraints feeds target terms into the source sentence with "source factors", which are extra per-token labels marking which words are glossary insertions. It reaches 93–94% term usage. Integrating Domain Terminology into Neural Machine Translation adds grammatical placeholder tags so injected terms can inflect. On hard terms this cuts absent-term errors from 34% (plain appending) to 2%. Lexically Constrained Neural Machine Translation with Levenshtein Transformer and Neural Machine Translation Decoding with Terminology Constraints force terms during decoding, the first with 100% usage. We keep their input formats and decoders. What we change is which constraints the model is trained on. We also add a "retired form" signal that none of them has.
Two papers motivate the gap. The agent-memory study above found that state drift accounts for up to 63.5% of confirmed errors and that chain-of-thought reasoning makes anchoring worse. Mitigating Temporal Misalignment by Discarding Outdated Facts shows that modelling how long facts last improves calibration. Both are monolingual and say nothing about bilingual lexicons. We borrow the agent-memory study's evaluation idea of separating "picked the superseded value" from other errors, and apply it to term pairs.
Three translation papers come closest to a temporal setting, all through COVID-19, and all handle it statically. On the Evaluation of Machine Translation for Terminology Consistency shows that BLEU, a word-overlap score, rewards systems that dump terms at sentence ends. It proposes TERM, a translation error rate that penalises terminology mistakes more heavily, and Window Overlap, which checks the words around each term. We use both, but they cannot detect a retired term that appears instead of, or alongside, the current one. TICO-19 and the Huawei WMT21 biomedical submission handle emerging terminology as a one-time snapshot.
AURORA updates monolingual retrieval indexes for new terms, and its lexical drift was synthetic. It does not touch translation decoding.
Research Program
The first component is a supersession test set. Each item is a source sentence whose term has a time-stamped history: an old target equivalent, a current one, and a date of change. We will draw items from two places. One is real renamings in the COVID-era biomedical material of TICO-19 and the WMT biomedical tasks. The other is official term revisions in the 75,000-pair CSTT Indic term bank used by LexGen, wherever revision history can be recovered.
Real supersessions will be few. We will therefore add controlled counterfactual supersessions on IATE/Wiktionary En–De terms and on CSTT terms. In these, a term's target is replaced by a plausible alternative, either a listed synonym or a sanctioned variant. We select terms spanning several orders of magnitude of training-corpus frequency for the old target form. Each item is labelled by its case: a new source term with an unknown target, which is the classic out-of-vocabulary case, or an unchanged source term with a changed target, which is the pure obsolescence case.
The second component is the method, which has two parts. The first is prior-contradicting constraint training. Source-factor training normally samples constraints from glossary matches in the training bitext. Instead, with some probability we sample a frequent term from that bitext, invent a replacement target form, rewrite the reference to use it with the correct inflections, and present the replacement as the constraint. The model thereby learns that glossary input overrides even its strongest associations. The second part is a third factor value marking a supplied retired form. A lightweight decoding penalty is applied to the token sequences of any inflected variant of that form. This is a negative constraint, which existing systems lack.
The third component is analysis. We will plot old-term leakage against the log frequency of the old form in training, separately for each method and for each case type. We will also check whether leakage concentrates in document positions after the first mention, or in inflected forms that the glossary does not list.
The primary metric is leakage rate: the fraction of items where any inflected form of the retired term appears in the output, matched by lemma. We report it together with current-term usage, TERM, Window Overlap, and ChrF++, a character-level overlap score. A human check of grammatical agreement on a sample guards against the "cheating" behaviour Alam et al. describe.
Baselines
The naive reference is an unconstrained Transformer translation model trained on the same bitext. For En–De this is the WMT17 setup. For English–Indic it is the Samanantar-trained base model of LexGen, which reports average ChrF++ of 61.05 in Biotechnology against LexGen's 64.94. This reference measures how strong the prior toward retired terms is.
The strongest positive-constraint baseline is source-factor training in the replace mode of Dinu et al.. The paper reports 93–94% term usage against a 76% unconstrained baseline on WMT newstest. We add the morphosyntactic placeholder variant of Michon et al., which reports a 2% absent-term error rate. That figure makes it the strongest choice for inflected targets.
The hard-enforcement baseline is the Levenshtein Transformer with constraint masking from Susanto et al.. It reports 100% usage and up to +1.16 BLEU. We include it because it guarantees the current term appears, which isolates whether leakage happens alongside the forced term.
We also run a strong in-context baseline: a general-purpose multilingual model prompted with the updated glossary entry and an explicit "X replaces Y" note. It tests whether the oracle-context anchoring reported for agent memory carries over to translation.
The corpus-rewrite baseline is described in the next section. No reported result exists on Lacuna for any method on terminology supersession or old-term leakage. All baseline numbers on our test set will be new.
Fairness rules apply to every system. All translation systems share the same training bitext, architecture size, and training steps. Our method's counterfactual examples replace ordinary constrained examples rather than adding extra data. Each method gets the same budget for tuning its constraint sampling rate.
Two ablations isolate the key ingredient. One trains with prior-contradicting constraints but without the retired-form factor. The other trains with standard sampling but with the retired-form factor. Beating the baselines means two things at once. Leakage must fall on high-frequency supersessions, which are the top frequency tercile. And current-term usage, ChrF++, and TERM must stay within noise of the best baseline on ordinary, non-superseded terms.
Why not just…?
Why not just update the glossary and keep using source-factor or placeholder training? This is the real threat. Michon et al.'s 2% error rate on hard terms suggests those methods may already resist strong priors. The training of both methods, however, never contrasts a constraint against a high-frequency competitor. Neither reports performance stratified by competitor frequency. The frequency-stratified leakage plot settles this directly. If source factors with an updated glossary show flat, low leakage across frequency, our training change is unnecessary. In that case we would report the benchmark and the negative finding instead.
Why not just force the new term with hard constrained decoding? Forcing guarantees the new string appears. It does not stop the model from also emitting the retired term, a duplication pattern Hasler et al. document for constrained search. It also breaks inflection, as both Dinu et al. and Susanto et al. note. We measure co-occurrence leakage and agreement errors for the Levenshtein baseline specifically.
Why not just rewrite the retired term throughout the training corpus and fine-tune? This is plausible whenever the old-to-new mapping is known. We run it as a baseline, rewriting inflected forms with a morphological analyser and fine-tuning with out-of-domain mixing as Huawei AARC recommends. Its weaknesses are cost and forgetting under repeated updates, and that study documents under-translation after small in-domain fine-tuning. We test it in a sequential setting of ten successive supersession batches, tracking quality on earlier terms. If corpus rewriting matches our method on leakage and forgetting at acceptable cost, that is a legitimate outcome, and the case for our method narrows to the setting where updates are too frequent to retrain for each one.
Why not just prompt a large model with the glossary and ask it to reason? The agent-memory study found that reasoning traces lowered accuracy on superseded facts. LexGen found that large models fall into transliteration, spelling the English term phonetically in the target script instead of using the native technical term. Our in-context baseline, run with and without reasoning, tests whether both effects hold for term supersession.
Decisive Checks
Three results decide the proposal. The first test runs on the strongest positive-constraint baseline: if its leakage does not rise with old-term frequency, the core premise fails. The second test is the prior-contradicting training ablation, which must reduce high-frequency leakage on real supersessions, not only on counterfactual ones; otherwise the method is learning the benchmark's artefact. The third test is the sequential-update experiment, which must show that our model absorbs successive batches of supersessions with less loss on earlier terms than corpus-rewrite fine-tuning. A secondary output is a first diachronic measurement on the test set: how term accuracy for current terms decays with the gap between a model's training cutoff and each term's change date. No existing benchmark in the file supports this measurement.
References
- Can Agent Memory Systems Track Evolving State?
- Training Neural Machine Translation to Apply Terminology Constraints
- Integrating Domain Terminology into Neural Machine Translation
- Lexically Constrained Neural Machine Translation with Levenshtein Transformer
- Neural Machine Translation Decoding with Terminology Constraints
- Mitigating Temporal Misalignment by Discarding Outdated Facts
- On the Evaluation of Machine Translation for Terminology Consistency
- TICO-19: the Translation Initiative for COvid-19
- Huawei AARC's Submissions to the WMT21 Biomedical Translation Task: Domain Adaption from a Practical Perspective
- AURORA: Neuro-Symbolic Continual Indexing for Evolving RAG Systems
- LexGen: Domain-aware Multilingual Lexicon Generation