---
title: 'Gender-Neutral Rewriting: Methods & Metrics'
url: https://www.emergentmind.com/topics/gender-neutral-rewriting-gnr
type: topic
---

# Gender-Neutral Rewriting: Methods & Metrics

Gender-neutral rewriting (GNR) is a text transformation task in which a sentence containing gendered references to human entities is reformulated into a gender-neutral equivalent while preserving meaning. In a canonical English formulation, the input is a gendered sentence referring to one human entity and the output is the same sentence rewritten into a gender-neutral version, typically using singular *they* forms; the task explicitly excludes sentences with more than one human entity because that would introduce ambiguity about which entity should be rewritten [2102.06788]. More generally, GNR is also described as reformulating a sentence to remove explicit gender markings referring to human entities, without altering the sentence beyond what is necessary for neutralization and ensuring semantic equivalence to the input [2509.13480]. The field now spans monolingual rewriting, post-editing of machine translation, and multilingual benchmarking, while remaining distinct from adjacent tasks such as Arabic user-centric gender rewriting, whose goal is to realize masculine or feminine forms for a specified speaker–listener context rather than neutralization [2205.02211].

## 1. Definition, scope, and task boundaries

GNR is not a gender-inference task. The English case study that formalized the task states that the goal is not to infer someone’s gender, but to fluently convert between forms when the preferred pronoun or style is known [2102.06788]. The same ethical stance recurs in later work: systems should not infer pronouns and should not prescribe a single “correct” gender form, but should support user-chosen identity expression [2102.06788].

The scope of neutralization is conditional rather than universal. In gender-neutral translation, a closely related formulation, neutralization is required when gender is unknown, unspecified, or should not be inferred, but gender should be preserved when it is clearly encoded in the source [2301.10075]. GeNTE operationalizes this distinction with **Set-N**, where English source sentences are gender-ambiguous or gender-irrelevant and should be rendered neutrally in Italian, and **Set-G**, where gender is explicit and should be translated with the correct masculine or feminine marking [2310.05294]. The CALAMITA GFG challenge adopts the same principle for Italian fair reformulation and cross-lingual generation: neutralize when the source does not warrant explicit gender marking, but do not over-neutralize when gender is semantically necessary [2412.19168].

A frequent misconception is that GNR consists only of pronoun replacement. The literature consistently treats it as broader controlled rewriting. In English it includes pronouns, possessives, verb agreement, and lexical gendered terms such as *fireman* → *firefighter* and *spokesman* → *spokesperson* [2102.06788]. In grammatical-gender languages, neutralization often requires paraphrase, restructuring, collective nouns, impersonal constructions, or passive voice rather than local substitution [2301.10075].

## 2. Linguistic operations and rewriting strategies

In the English singular-*they* setting, the central operations are one-to-many pronoun mappings and subject–verb agreement. Pronouns such as `he` and `she` map to `they`, while `his`, `her`, `hers`, `him`, `himself`, and `herself` may map to different neutral forms depending on context; `he grows` becomes `they grow`, but verbs not governed by the gendered subject remain unchanged [2102.06788]. The main technical challenges identified in that work are precisely these one-to-many mappings and correctly identifying the relevant verbs [2102.06788].

Rule-based English systems make these distinctions explicit. NeuTral Rewriter maps `he, she → they`, `him → them`, `hers → theirs`, and resolves ambiguous forms such as `her` and `his` using Stanza POS tagging and dependency parsing; it then applies a LanguageTool grammar correction pass to fix agreement changes such as `works → work`, `has → have`, and `is → are` [2109.06105]. The same paper extends beyond pronouns to occupations, titles, and generic *man* constructions, including `chairman → chairperson`, `policeman → police officer`, `fireman → firefighter`, `stewardess → flight attendant`, `mankind → humankind`, and `freshman → first-year student` [2109.06105].

For Italian and other grammatical-gender languages, the strategy inventory is much wider. A guideline-based synthesis of English–Italian gender-neutral translation groups operations into epicene synonyms, pluralization, relative or indefinite pronouns, collective or role nouns, omission, repetition, passive voice, imperative forms, and impersonal forms [2301.10075]. Concrete examples include *doctors* → *personale medico* and *degli scrittori* → *di chi scrive* [2402.06041]. The Italian GNR study emphasizes that even a simple sentence can contain many gender-marked words, so acceptable neutral alternatives are often not one-to-one replacements [2509.13480].

French work has pushed neutralization toward collective-noun rewriting. GeNRe rewrites masculine plural human nouns into collective nouns whose grammatical gender is fixed and does not depend on the gender of the people referred to, for example `Les lecteurs assidus financent le journal chaque mois.` → `Le lectorat assidu finance le journal chaque mois.` [2505.23630]. This differs from inclusive orthographic strategies because it aims to avoid gender marking altogether by standard lexical means [2505.23630].

## 3. Corpora, benchmarks, and annotation practices

The English singular-*they* benchmark introduced in 2021 was built without human-labeled training data. The authors started from **100 million Wikipedia sentences**, filtered them for a single gendered entity, excluded sentences containing both masculine and feminine pronoun sets, and obtained **15 million** gendered sentences. They then created pseudo-parallel data with an automatic rewriting algorithm and evaluated on a human-curated test set of **500 sentence pairs**, drawn equally from Twitter, Reddit, news articles, movie quotes, and jokes, with equal numbers of masculine and feminine source sentences in each domain [2102.06788].

NeuTral Rewriter broadened English benchmarking with both synthetic and natural data. It introduced **WinoBias+** with **3,167 synthetic sentences** and manually curated neutral alternatives, and natural benchmarks of **1,000 manually corrected parallel sentences** each from OpenSubtitles and Reddit, balanced across target forms such as `he`, `she`, `her(s)`, `his`, and `him/herself` [2109.06105].

For Italian, GeNTE is a dedicated benchmark for gender-neutral translation and rewriting. It contains **1,500 English–Italian parallel sentence pairs**, split evenly into **Set-N** and **Set-G**. A **common-set** of **200 sentences** includes one original gendered reference and **3 independently created neutral references** per source sentence, making explicit the open-endedness of neutral rewriting; only **13.57%** of cases had identical neutral references across all three translators [2310.05294]. The multilingual extension mGeNTE generalizes this design to English–Italian, English–German, and English–Spanish, yielding **4,500 entries total** and a **PARALLEL-SET** of **1,106 sentences** that are fully parallel across all three language pairs; all target-side gendered words were annotated, with Dice agreement of **0.95** for en-de, **0.94** for en-es, and **0.95** for en-it [2501.09409].

Translation into English from weakly gendered languages has produced a separate benchmark family. GATE X-E defines the **Arbitrarily Gender-Marked Entity (AGME)** as an individual who is not gender-marked in the source but is gender-marked in the English translation through a gendered pronoun or intrinsically gendered noun. It provides human feminine, masculine, and neutral variants for Turkish, Persian, Finnish, and Hungarian source sentences, with **1,429** tr→en instances, **1,259** fa→en, **1,832** fi→en, and **1,308** hu→en [2402.14277].

## 4. Methods and modeling paradigms

A recurring pattern in GNR is a high-precision rule-based system used either directly or as a generator of pseudo-parallel training data. The 2021 English singular-*they* system combines regular expressions, a SpaCy dependency parser, and GPT-2. Regular expressions handle straightforward substitutions and lexical neutralization; the parser identifies which verbs are governed by the gendered subject and pluralizes them appropriately; GPT-2 resolves ambiguous pronouns by generating all candidate rewrites and selecting the one with lowest perplexity [2102.06788]. A Transformer seq2seq model with **6 encoder layers**, **6 decoder layers**, implemented in **Fairseq**, is then trained on the automatically generated pairs with per-token cross-entropy and augmented with identity pairs and gender-inflected sentences [2102.06788].

This pipeline achieved extremely strong English results. On the 500-pair human-curated test set, the identity baseline scored **BLEU 90.32** and **WER 12.40%**, the algorithm scored **BLEU 99.63** and **WER 0.63%**, and the learned Transformer scored **BLEU 99.44** and **WER 0.99%** [2102.06788]. NeuTral Rewriter follows a related recipe—rule-based rewriting plus a neural fairseq Transformer trained on synthetic Reddit pairs—and reports **WER 0.00** on WinoBias+, **0.18** on OpenSubtitles, and **0.02** on Reddit for the neural model [2109.06105].

Later work diversified the modeling space. For German, a gender-fair rewriting model reverses the usual pseudo-data direction: instead of starting from biased text and generating fair targets, it starts from real gender-fair text and creates pseudo-biased sources, first with backward augmentation for English and then with round-trip translation for German. On a German TIGER-based test set, the best round-trip configuration reached **WER 13.18**, improving on the biased source at **20.56** and on a rule-based baseline at **14.02** [2305.11140]. For Spanish conversational sentences, a Transformer NMT system trained on a **2.2M** gender-parallel corpus rewrites short utterances into opposite-gender variants and can be conditioned with source tags distinguishing re-genderable from neutral sentences [2107.05987].

Two newer lines of work move beyond standard seq2seq rewriting. DEPEN, “Detect and Perturb to Neutralize,” first masks tokens that most strongly predict a sensitive attribute using BERT self-attention salience, then regenerates with a BART model whose decoder hidden states are perturbed at inference time so that a sensitive classifier’s output distribution is pushed toward uniform uncertainty [2109.11708]. At a still more internal level, neuron-level interventions identify masculine, feminine, and gender-neutral neurons in Transformer MLP blocks and perform keep-only masking during generation; on curated ternary datasets, the method preserves or slightly increases target gender ratios while reducing leakage into non-target categories, although the neutral target is described as the hardest setting [2605.30717].

## 5. Multilingual and cross-lingual expansion

A major extension of GNR is gender-neutral translation (GNT), especially when translating from English into grammatical-gender languages. In this setting the task is “the task of automatically translating from one language into another without marking the gender of human referents in the target” [2301.10075]. Prompt-based experiments on English→Italian show that off-the-shelf systems are not adequate: Amazon Translate, DeepL, and GPT-4 without special prompting produce only about **3% neutral translations**, with about **97%** remaining gendered, mostly masculine. With prompting, GPT-4 produces roughly **65–70% fully neutral outputs** plus about **15% partially neutral outputs**, though acceptability remains variable [2402.06041].

Evaluation work on Italian GNT has shown that neutrality is an open-ended generation problem. On GeNTE, reference-based metrics such as BLEU, chrF, TER, METEOR, BERTScore, COMET, and BLEURT do not reliably capture neutralization, particularly in **Set-N**, because valid outputs often use paraphrase, synonymy, or structural rewording. A reference-free UmBERTo classifier trained on about **380,000 sentences** of synthetic neutral, masculine, and feminine Italian data yields more balanced accuracy, with **89.83** overall across DeepL outputs on Set-G and Set-N combined [2310.05294].

Translation into English from weakly gendered languages exposes a different configuration. On the pronoun-only subset of GATE X-E, a GPT-3.5 Turbo neutral rewriter prompted with “Change all gendered pronouns to use singular ‘they’ instead” reaches very high neutral rewriting accuracy, including **98.90** for tr→en few-shot, **97.00** for hu→en few-shot, **96.77** for fi→en few-shot, and **97.84** for fa→en few-shot; once a reliable neutral rewrite exists, deterministic rules generate feminine and masculine alternatives with roughly **98–99.9% accuracy** [2311.08836]. A GPT-4 chain-of-thought system evaluated on GATE X-E performs very well on pronoun-only rewriting, with accuracy roughly **0.96–0.99**, but much worse on gendered-noun cases, where exact-match accuracy is generally around **0.5–0.8** [2402.14277].

Not all multilingual work belongs to GNR in the strict sense. Arabic user-centric gender rewriting models four target contexts—**1M/2M**, **1F/2M**, **1M/2F**, **1F/2F**—for first- and second-person discourse participants in Modern Standard Arabic, and the shared task built on APGC v2.1 evaluates these systems with M² and BLEU rather than neutralization metrics [2205.02211]. This adjacency is important: it situates GNR within a larger family of gender-aware controlled rewriting, but with a different target semantics.

## 6. Evaluation criteria, error profiles, and open issues

Evaluation in GNR is fundamentally multi-objective. English work commonly reports **BLEU** and **Word Error Rate (WER)** [2102.06788], while Italian work has argued for a two-dimensional framework separating **neutrality** from **semantic fidelity / meaning preservation** [2509.13480]. In that framework, neutrality is measured with an LLM-as-a-Judge binary decision, and meaning preservation is measured with BERTScore. Using mGeNTE human references, the Italian study derives a conservative human-level BERTScore threshold of **0.879** from a mean of `0.9334` and a standard deviation of `0.0546` [2509.13480].

Error analyses across languages show recurrent failure modes. In English singular-*they* rewriting, rule-based errors come from dependency parser mistakes and GPT-2 perplexity ranking of ambiguous pronouns, while seq2seq model errors are often about rare tokens, whitespace, emojis, and symbols such as `@`, `%`, and `*`, especially because training data are drawn from relatively clean Wikipedia but testing includes Twitter and Reddit [2102.06788]. Both systems make more mistakes on feminine inputs than masculine ones, and the authors attribute this to Wikipedia training data being about **70% masculine** [2102.06788]. In French collective-noun rewriting, the dominant error class is morphosyntax—especially verbs and adjectives—because rewriting from plural member nouns to singular collective nouns changes agreement throughout the clause [2505.23630].

A second broad issue is subjectivity. Manual evaluation of English→Italian GNT uses **N**, **G**, and **P** labels for fully neutral, fully gendered, and partially neutral outputs, plus a four-point acceptability scale. Across **1,800** outputs judged by three annotators, Fleiss’ kappa for neutrality is **0.89**, but ICC for acceptability is only **0.48**, and the study explicitly notes that acceptability is subjective and does not have a single correct answer [2402.06041]. The existence of multiple acceptable neutral references in GeNTE and mGeNTE reinforces the same point [2310.05294].

A third issue is semantic trade-off. DEPEN shows that stronger neutralization can sacrifice fidelity, while rule-based systems often preserve the input too closely and leave bias behind [2109.11708]. The Italian GNR study finds an analogous trade-off between optimizing training data for neutrality and for meaning preservation: training on the full dataset yields the largest neutrality gains, while training on the clean BERTScore-filtered subset better preserves meaning but produces smaller neutrality gains and can even reduce neutrality for some models [2509.13480].

## 7. Applications, ethics, and research directions

The practical applications of GNR recur across the literature. The original English work argues that if a translation system or writing assistant encounters gendered language, it should be able to produce a neutral alternative, and also highlights data augmentation for robustness as a use case [2102.06788]. Arabic user-centric rewriting is demonstrated as a post-editing wrapper around Google Translate to personalize outputs according to speaker and listener gender preferences [2205.02211]. GEC work shows that targeted augmentation for singular *they* reduces explicit bias cases from **32** to **8** or from **34** to **7** in GECToR variants while preserving overall correction quality, indicating that gender-aware rewriting techniques can serve as bias mitigation for correction systems as well [2306.07415].

Ethically, the literature is explicit that neutralization should support, not override, identity. The English singular-*they* task states that systems should not infer pronouns and should instead support user-chosen identity expression [2102.06788]. DEPEN adds that neutralization systems can be misused or can inadvertently strip away legitimate information, and therefore recommends human auditing rather than black-box deployment [2109.11708]. Translation work similarly stresses that GNR should distinguish cases where gender is unknown from cases where it is semantically relevant, rather than treating all gender marking as an error [2301.10075].

Current directions indicate both broader coverage and deeper control. mGeNTE provides multilingual aligned resources for Italian, German, and Spanish neutral rewriting and language modeling [2501.09409]. The GFG challenge extends evaluation to monolingual Italian reformulation, cross-lingual generation, and neomorpheme-based rewriting through Neo-GATE [2412.19168]. Neuron-level intervention work suggests that gender-neutral generation is a distinct controllable behavior inside language models, not merely the absence of masculine or feminine markers, although neutral control remains more fragile than binary gender switching [2605.30717]. A plausible implication is that future GNR systems will combine explicit linguistic constraints, richer multilingual benchmarks, and internal-control methods while continuing to balance neutrality, fluency, and adequacy rather than optimizing any single objective [2402.06041].

Source: https://www.emergentmind.com/topics/gender-neutral-rewriting-gnr