Businesses, hospitals and regulators increasingly force machine translation systems to use approved vocabulary. They supply a term base, a curated bilingual glossary, and the translator must use its entries. These entries are almost always stored in dictionary citation form. For person nouns in German, Italian, Spanish or French, that citation form is usually masculine: Ingenieur, chirurgo, médico. Separately, a large body of work shows that translation systems already over-produce masculine forms even when the English source says the person is a woman.
Our starting hypothesis is that these two facts compound. When a system is told "use the lemma Ingenieur here" (a lemma is the base dictionary form), that instruction adds a second masculine anchor on top of the statistical one. Terminology control could therefore make gender errors worse, precisely in the regulated domains where both terminology and correct reference to people matter.
We propose to test this "citation-form anchoring" effect directly. If it holds, we will replace single-lemma constraints with gender-complete term entries. Each approved concept would list its masculine, feminine and, where one exists, epicene form (a single form used for any gender, like Italian docente). A lightweight source-side resolver would decide which form the context licenses. The translation model would be trained to propagate that choice through the articles, adjectives and participles that must agree with the noun.
Core Claim
The claim has two parts, and the first is the one that justifies the project. First, terminology-constrained systems of the copy-and-inflect family will show lower feminine accuracy and a larger masculine–feminine gap than the same system without constraints, whenever the injected lemma is masculine. Second, this damage can be removed without losing terminology adherence. The fix treats the approved concept as fixed and the gendered surface form as a context-dependent choice.
This is the decoupling that translation studies call hybrid strategy: preserve the conceptual layer, adapt the grammatical layer. Yin et al. do this for poetry, translating imagery literally while conforming syntax to the target language. Here the "concept" is the standardized term, and the "structure" is its gender inflection and agreement cascade.
When the source gives no gender cue at all, the same logic yields a third behaviour. The system keeps the approved concept but uses its epicene variant or an impersonal construction. This follows the dynamic constraint of Gender Neutralization for an Inclusive Machine Translation, which argues that systems should neutralize when gender is unknown and preserve it when specified. That paper builds no system. We treat the neutral case as a secondary, human-evaluated component, because parallel data for neutral translation is scarce, as that paper itself stresses.
Why This Gap Is Real
Three lines of terminology work supply the machinery we reuse. All three ignore social gender.
Training Neural Machine Translation to Apply Terminology Constraints appends target terms to the source, marked by extra input features, so the model learns to copy and inflect them. Facilitating Terminology Translation with Target Lemma Annotations shows that annotating with lemmas rather than exact forms works far better in morphologically rich languages. Morphology Aware Source Term Masking replaces the source term with the target lemma plus morphological tags extracted from the source.
The lemma papers evaluate case and number in automotive and news text, never person nouns with feminine counterparts. Morph-masking reads gender off Basque source morphology. It has nothing to read when the source is the English word "doctor," whose gender lives in a pronoun or a word like "pregnant" elsewhere in the sentence. Rule-based Morphological Inflection Improves Neural Terminology Translation inflects the term with dependency-parse rules. It models syntactic case, not referent gender, and it inflects the term alone rather than its dependents.
The gender literature motivates the hypothesis. Investigating Failures of Automatic Translation in the Case of Unambiguous Gender shows that systems including M2M-100 stay below 70% accuracy even when an explicit cue fixes the gender, with a consistent penalty for female referents. A Tale of Pronouns attributes such errors to the model assigning near-zero attribution to the gender-marking pronoun. Identifying Gender Stereotypes and Biases in Automated Translation From English to Italian Using Similarity Networks finds that neutral terms drift toward masculine anchors in the target representation space.
If the model already under-weights the feminine cue, handing it an explicit masculine lemma should tip it further. Nothing in the file measures this. We found no reported result on Lacuna for any terminology-constrained system on WinoMT, MuST-SHE or the unambiguous-gender benchmark.
Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphology constrains the design and also supplies a tool. It shows that swapping gender on the noun alone produces ungrammatical text. It then propagates the change through the dependency tree with a Markov random field, a probabilistic model over the parse. We reuse that propagation to build training data, which it never used for bilingual terminology translation.
Research Program
The work has three stages.
Diagnosis. We build a gendered-terminology test suite for English into German, Italian and Spanish. We start from the WinoMT sentences of Evaluating Gender Bias in Machine Translation, MuST-SHE's unambiguous category from Test Suites Task: Evaluation of Gender Fairness in MT with MuST-SHE and INES, and person-noun sentences from the medical TICO-19 set (TICO-19). We attach IATE-style glossary entries to every occupation or role noun, in two variants. In the realistic variant, the entry is the masculine citation lemma, as term bases typically store it. In the gender-complete variant, the entry lists the masculine, feminine and, where attested, epicene forms.
We then train lemma-annotation and append-style systems exactly as in the Dinu and Bergmanis setups. We measure feminine accuracy, masculine accuracy and their gap, with and without the realistic constraint.
We define the anchoring effect as the drop in feminine accuracy caused by adding the masculine-lemma constraint to the same model on the same sentences. It is the quantity the whole project hinges on. As a side product, this stage also audits how often existing glossary resources lack a feminine or epicene form for person nouns. That audit is the standardization finding: a term base that stores one masculine lemma per concept is itself a source of bias.
Method. The source side receives the approved concept as a term family, plus a gender tag with three values: feminine, masculine or unknown. A resolver produces the tag. It combines an off-the-shelf English coreference model with a small lexicon of gender triggers such as pronouns, kinship nouns and "pregnant," and abstains to "unknown" below a confidence threshold. This is the step morph-masking lacks for English sources: it moves a non-local cue onto the term slot.
For training, we take parallel sentences containing glossary person nouns. We create gender-flipped counterparts of the target side using the dependency-level agreement propagation of Zmigrod et al., and pair each version with the matching tag. The model therefore sees the same concept constraint with both gender values, and must realize the whole agreement cascade to match.
For the unknown value, the target is the epicene entry where one exists. Otherwise it is a structurally neutralized rewrite. We produce these rewrites with a large language model and have native speakers verify them on a sample, keeping only verified patterns for evaluation claims.
Evaluation. Standard string metrics punish legitimate neutral rewrites. Following Piergentili et al.'s warning, we therefore report reference-free gender accuracy on the term, dependent-agreement accuracy, and concept-level term accuracy. Concept-level accuracy counts any member of the approved family as correct and any unapproved synonym as wrong. We report COMET and chrF on general test sets as a check that general quality is unchanged. Neutral outputs are judged by humans.
Baselines
The naive reference point is the same Transformer trained on the same parallel data with no terminology input. It gives the unconstrained gender accuracy that anchoring is measured against, and the low term accuracy any constrained method must beat.
The strongest existing terminology method is target lemma annotation from Facilitating Terminology Translation with Target Lemma Annotations. It reports 93% term accuracy in English–Latvian human evaluation, against 55% for the unconstrained baseline. We rerun it on our data with the realistic masculine-lemma glossary, sharing architecture, data and tuning budget with our system. To beat it, we must match its concept-level term accuracy within two points while clearly improving feminine accuracy and shrinking the gender gap.
We also run morph-masking from Morphology Aware Source Term Masking in two forms. In one, the gender tag comes from a source tagger, which for English yields nothing. In the other, we graft our resolver's tag onto it. This separates the value of resolving gender from the value of our training data.
The strongest gender-debiasing baseline is interpretability-guided few-shot prompting from A Tale of Pronouns. On English–Spanish WinoMT with Flan-T5, it reduced the gender gap from 7.2 to 2.1 and raised accuracy by 7.1 points. We run it with the glossary added to the prompt, on the same test suite, with the same demonstration budget. For reference ceilings on general gender translation, MuST-SHE reports 80.07% overall gender accuracy for the best English–German system but 50% on ambiguous feminine first-person cases. WinoMT reports a French commercial system at 80% on stereotypical roles and 54% on anti-stereotypical ones.
Three ablations isolate the key ingredients. We replace the resolver's tag with the oracle gender to bound resolver error. We remove the counterfactual gender-flipped training data while keeping the tag, which tests whether the model obeys a gender feature it has rarely seen contradicted. We supply the gender-complete family without any tag, which tests whether listing the feminine form alone removes anchoring.
Why not just…?
Why not just put the feminine lemma in the glossary constraint when the context is feminine? This is the most serious alternative, and it is the "gender-complete family, no counterfactual training" ablation plus the resolver. If lemma annotation already inflects dependents correctly once given Ingenieurin, most of our method reduces to the resolver and a better term base. That would still be a useful standardization result, but a smaller one. The dependent-agreement metric settles this. The Zmigrod et al. finding that agreement must be propagated suggests the model will often produce the feminine noun with masculine articles or adjectives. We will report the outcome either way.
Why not just inflect the term with rules after translation? Rule-based Morphological Inflection Improves Neural Terminology Translation reaches 87.6% inflection accuracy in German, where a neural inflection module reached only 15.4% in Lithuanian. That is strong evidence that rules help with rare terms. However, it changes only the term. Changing gender also requires rewriting agreeing words, and noun-only swaps produce ungrammatical text according to Counterfactual Data Augmentation. We run this pipeline with the resolver's gender feature added, and compare dependent-agreement accuracy directly.
Why not just prompt a large language model with the glossary and an instruction to avoid masculine defaults? This is a real threat for high-resource pairs. GPT-4 with five examples scored best on INES, but only 65.71% inclusivity accuracy according to Test Suites Task. MiTTenS shows worst-case collapses as low as 14.3% for Mistral 7B. Neither result tests glossary-constrained prompting. We therefore include a strong prompted model with glossary and instruction as a baseline, and report it separately on the in-domain medical subset. If it matches our system, the contribution shrinks to the diagnosis and the term-base schema, and we will say so.
Proposed Experiments and Decisive Checks
The first decisive check is the anchoring effect itself. The unconstrained and lemma-constrained systems see identical English–German, English–Italian and English–Spanish sentences, and we compare feminine accuracy on WinoMT anti-stereotypical and MuST-SHE unambiguous items. A drop of several points, significant under paired bootstrap, confirms the premise. If there is no drop, the project reduces to an inclusive term-base proposal and we would stop the method work.
The second check tests whether gender-complete entries with counterfactual training close the gap to the unconstrained system or better. On the realistic subset, concept-level term accuracy must stay within two points of target lemma annotation. On general test sets, COMET must stay within noise of the unconstrained system.
The third check measures dependent-agreement accuracy against the rule-based and tag-only variants. The claim is that training on propagated counterfactuals, not merely supplying the right noun, fixes the cascade.
The fourth check is resolver robustness. We run the 20-language unambiguous-gender benchmark's German, Italian and Spanish portions, measuring the oracle–resolver gap to see how much remaining error lies upstream.
Finally, on ambiguous items, native-speaker raters judge whether "unknown" outputs keep the approved concept, avoid gendered marking, and remain grammatical. This is reported as exploratory, given the limited neutral training data.
References
- Integrated Adoption of Domestication and Foreignization: A Study on the English Translation of Prelude to the Melody of Water by Su Shi
- Gender Neutralization for an Inclusive Machine Translation: from Theoretical Foundations to Open Challenges
- Training Neural Machine Translation to Apply Terminology Constraints
- Facilitating Terminology Translation with Target Lemma Annotations
- Morphology Aware Source Term Masking for Terminology-Constrained NMT
- Rule-based Morphological Inflection Improves Neural Terminology Translation
- Investigating Failures of Automatic Translationin the Case of Unambiguous Gender
- A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation
- Identifying Gender Stereotypes and Biases in Automated Translation From English to Italian Using Similarity Networks
- Counterfactual Data Augmentation for Mitigating Gender Stereotypes in Languages with Rich Morphology
- Evaluating Gender Bias in Machine Translation
- Test Suites Task: Evaluation of Gender Fairness in MT with MuST-SHE and INES
- TICO-19: the Translation Initiative for COvid-19
- MiTTenS: A Dataset for Evaluating Gender Mistranslation