When a translation system turns an English conversation into Italian, German or Spanish, it often has to pick a grammatical gender for a word like "doctor" or "librarian." The only evidence may sit several turns earlier: "she" in a previous message, or a name. Current systems mostly fail at this. On inter-sentential cases, commercial translators land near coin-flip accuracy, and giving a model more context can make things worse, because the extra text also carries domain signals such as parliamentary language that push the model toward masculine forms. We propose to split the problem in two. First, an explicit rewriting step, of the kind used in conversational search to turn "Where does she practice?" into a standalone question, carries the referent's gender into the current turn. Second, a sentence-level translator receives only that rewritten turn and no raw history.
The contribution is diagnostic as much as it is a method. Explicit rewriting creates an inspectable intermediate text, so masculine-default errors can be attributed to one of two stages. Either the system failed to recover who a phrase refers to, or it recovered the gender and the translator overrode it. It also tests a sharp prediction. If context hurts because it leaks domain priors, then passing only the resolved referent should remove that harm while keeping the benefit of the gender cue.
Core Claim
Masculine defaults in context-dependent translation mix two separable failures. The first is resolution: the antecedent's gender is not carried to the current mention. The second is realization: a gender cue inside the sentence is ignored in favor of the occupational stereotype. Raw concatenation of history confounds the two and adds a third effect, domain-prior amplification. The claim that separates this proposal from existing work is that a gender-preserving rewriter placed in front of a sentence-level translator does two things. It removes the domain-amplification effect found by Does Context Help Mitigate Gender Bias in Neural Machine Translation?, which showed that 2-to-1 concatenation raises English-German feminine accuracy from about 5% to 23% but lowers it when the context comes from masculine-skewed domains. And it exposes, as a measurable residual, how much of the remaining error is pure realization failure.
We state the decomposition explicitly because it organizes every experiment. Let $R$ be the event that the rewritten turn carries the correct referent gender. Let $C$ be the event that the translation inflects correctly. Then
$$P(C) = P(R)\,P(C \mid R) + \big(1-P(R)\big)\,P(C \mid \neg R).$$
The first factor measures the rewriter. The second, $P(C \mid R)$, measures the translator's willingness to obey an explicit cue. It can be estimated directly by feeding gold rewrites.
Why This Is Not Already Done
The closest work treats context as undifferentiated input. MT-GenEval built a contextual subset of Wikipedia sentence pairs whose gender is recoverable only from the preceding sentence. It reports commercial accuracy of 50–61% against over 95% for professional translators, and finds that gender-filtered self-training combined with contextual training mitigates best. It motivates our work but uses two-sentence prose, not dialogue, and it never separates resolution from realization. Gete and Etchegoyhen constrain our design: their domain finding is why we refuse to pass raw history to the translator. Recent Advances in Dialogue Machine Translation surveys dialogue translation, where cross-turn coreference and zero pronouns are central, but it evaluates cohesion and BLEU and never measures gender. The conversational query rewriting line supplies the method. Scaling Multi-Domain Dialogue State Tracking via Query Reformulation showed that rewriting lets single-turn components run over multi-turn input unchanged, and ConvGQR and AdaCQR are strong trained rewriters. All of them optimize retrieval metrics, and none checks whether "they" or "she" survives the rewrite. So we also audit these rewriters for masculine injection, a failure we have independent reason to expect. Adapting Psycholinguistic Research for LLMs found that English models turn singular "they" into masculine pronouns and that German models favor "men" regardless of antecedent.
One paper threatens the proposal directly. Investigating Failures of Automatic Translation in the Case of Unambiguous Gender shows that systems fail even when the cue is inside the sentence, as in "My doctor is pregnant." If that failure dominates, rewriting only moves the problem from resolution to realization. We treat this as a hypothesis to measure, not an objection to dismiss. The estimate of $P(C \mid R)$ says how much of the problem rewriting can ever solve, and a low value is itself a publishable finding about where effort should go.
Research Program
The first component is a conversational gender-resolution test set. It is built from the MT-GenEval contextual items and from COH-TGT:GENDER for Basque-Spanish. We turn each two-sentence item into a short dialogue by inserting a controlled number of intervening turns between the gender-bearing antecedent and the ambiguous target mention. Three factors vary. The first is distance, from zero to four intervening turns. The second is a topic switch modeled on the 3–4-turn topic segments reported in TopiOCQA, which introduces a competing person of the other gender. The third is the domain register of the filler turns, masculine-skewed parliamentary text versus TED-style talk, reproducing the Gete and Etchegoyhen manipulation in dialogue form. Each item has a gender-flipped counterfactual twin and a stereotype label (pro- or anti-stereotypical occupation), so stereotype boost can be measured directly. Filler turns are machine-generated and then human-checked so that they contain no competing gender cues except where the design intends one.
The second component is a gender-preserving rewriter. We start from T5-style rewriters trained on QReCC, as in ConvGQR and AdaCQR, and from a prompted open LLM. We then add a reference-free training signal in the spirit of Breaking the Reference Bottleneck. The reward is whether the rewrite keeps the antecedent's gender marker (pronoun, name or explicit phrase) attached to the target mention, and whether it introduces no gendered token absent from the history. We apply the reward through preference optimization, as PA-CQR did for entity omission. The rewriter is told to copy the referent, not the surrounding domain text. This design choice is what should cut off domain priors, and an ablation below tests it.
The third component is analysis. We estimate each term of the decomposition per language, distance, stereotype condition and domain. We also run the attribution method of A Tale of Pronouns on the translator for cases where the gold rewrite was supplied and the output still defaulted to masculine. The question is whether realization failures show the same low attribution to gender tokens that paper reported.
Baselines
The naive reference point is a sentence-level translator that sees only the current turn. With no cue available, its feminine accuracy on ambiguous targets measures the raw masculine prior. Gete and Etchegoyhen report about 5% feminine generative accuracy for English-German sentence-level models on MT-GenEval.
The strongest simple competitor is context concatenation. The translator receives the previous turn (2-to-1, as in Does Context Help Mitigate Gender Bias in Neural Machine Translation?) or the full dialogue history. It is the obvious answer, and it is the reported result we must beat: feminine generative accuracy of about 23% in English-German. For fairness we use the same translation backbone and training data for concatenation and for our pipeline, and the same tuning budget for every system. Beating it means higher feminine accuracy overall, a smaller masculine-feminine accuracy gap, and above all no drop under masculine-domain filler where concatenation drops.
The strongest existing mitigation is gender-filtered self-training plus contextual training from MT-GenEval, retrained on the same backbone. The file gives no numeric accuracy for it, only that it was the strongest combination. The paper's reported reference points are commercial contextual accuracy of 50–61% and human accuracy above 95%. We also run a full-history instruction-tuned LLM translator prompted with the whole dialogue. We found no reported result on Lacuna for LLM translators on the MT-GenEval contextual subset, so we will establish that number ourselves.
Two ablations isolate the key ingredient. The first, a "leaky rewrite" ablation, uses the same rewriter but appends the domain-bearing filler text to the rewritten turn. If the benefit disappears under masculine-domain filler, the gain comes from blocking domain priors and not from rewriting as such. The second, an oracle-rewrite condition, supplies a human-written gender-explicit turn. It gives $P(C \mid R)$ and the upper bound of the approach. A final condition places an off-the-shelf retrieval-tuned rewriter (ConvGQR- or AdaCQR-style, trained only for retrieval) in the pipeline. It tests whether gender-aware training of the rewriter matters.
Why not just…?
Why not just concatenate the dialogue history and let the model sort it out? Gete and Etchegoyhen already show that concatenation helps mainly on stereotypically feminine occupations and hurts under masculine-skewed context. Longer dialogue history gives more room for that leakage. Our domain-register factor tests this directly. If concatenation's feminine accuracy stays flat across parliamentary and TED filler in dialogue form, the domain-leak explanation fails and our pipeline's main advantage goes with it. We would report that.
Why not just prompt a strong LLM with the full conversation? This is a real threat, and there is no reported number to dismiss it. The evidence against it is indirect. A Tale of Pronouns finds that instruction-tuned translators give little attribution to gender-marking pronouns, and the psycholinguistic study finds masculine defaults in coreference itself. We run this system as a baseline across distance and topic-switch conditions. If it matches the pipeline, the contribution narrows to the diagnostic decomposition. That result would still show whether frontier models fail at resolution or at realization.
Why not just fine-tune the translator on contextual, gender-balanced data? MT-GenEval's self-training plus contextual training is exactly this, and we retrain it on the same backbone. It may close the gap at distance one. Our prediction is that it degrades with distance and topic switches, because it still learns from raw adjacent context. If it does not degrade, fine-tuning is the better choice and rewriting is only an interpretability tool.
Why not just run a coreference tagger and insert a gender tag instead of rewriting? For translation alone this may work as well, and the core claim does not depend on the rewriter's form. It depends on an explicit bottleneck that carries only the referent. We include a tag-insertion variant built from the same resolver outputs. The rewriter earns its place only if it handles ellipsis and topic-switch cases that tagging misses, or if the same rewrite also serves conversational search downstream.
Proposed Experiments
The first decisive check is the domain-register interaction. For concatenation, feminine accuracy should fall under masculine-domain filler relative to TED-style filler. For the rewrite pipeline, the difference should be near zero, and the leaky-rewrite ablation should reintroduce the drop. This reproduces and then removes the Gete and Etchegoyhen effect in dialogue.
The second is the decomposition itself. For each language pair (English into German, Italian, French and Spanish from MT-GenEval, plus Basque-Spanish) we report $P(R)$, $P(C \mid R)$ from oracle rewrites, and their split by stereotype condition. The finding Investigating Failures of Automatic Translation in the Case of Unambiguous Gender leads us to expect is that $P(C \mid R)$ stays well below one for anti-stereotypical items. That would locate the remaining masculine default in the translator, not in context handling.
The third is the rewriter audit. On the constructed dialogues, and on gendered-person subsets drawn from QReCC and TopiOCQA conversations, we measure how often retrieval-tuned rewriters drop, flip or masculinize a pronoun, especially singular "they." We also check whether gender-aware preference training lowers that rate without costing retrieval MRR. The reference points are ConvGQR's 42.0 MRR on QReCC dense retrieval and AdaCQR's 52.4 MRR under BM25.
For all systems we report gender accuracy with counterfactual twins, the feminine-masculine accuracy gap, and reference-based quality split by gender, following MT-GenEval's quality-gap analysis. We avoid learned quality estimators for model selection, because Watching the Watchers shows they penalize feminine outputs and would bias our comparison toward masculine defaults.
References
- Does Context Help Mitigate Gender Bias in Neural Machine Translation?
- MT-GenEval: A Counterfactual and Contextual Dataset for Evaluating Gender Accuracy in Machine Translation
- Recent Advances in Dialogue Machine Translation
- Scaling Multi-Domain Dialogue State Tracking via Query Reformulation
- ConvGQR: Generative Query Reformulation for Conversational Search
- AdaCQR: Enhancing Query Reformulation for Conversational Search via Sparse and Dense Retrieval Alignment
- Adapting Psycholinguistic Research for LLMs: Gender-inclusive Language in a Coreference Context
- Investigating Failures of Automatic Translationin the Case of Unambiguous Gender
- TopiOCQA: Open-domain Conversational Question Answering with Topic Switching
- Breaking the Reference Bottleneck via Learning to Rewrite Conversational Queries without Gold Reference Passages
- Improving Contextual Query Rewrite for Conversational AI Agents through User-preference Feedback Learning
- A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation
- Watching the Watchers: Exposing Gender Disparities in Machine Translation Quality Estimation