Source-linked AI summary
Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model
Chantal Amrhein, Florian Schottmann, Rico Sennrich, Samuel Läubli
TL;DR
Gender bias persists in generated language, and rule-based rewriting-data creation is difficult for morphologically complex languages. The paper reverses augmentation and uses biased machine translation in round trips to train a German rewriter. It reports improved German WER against a strong rule-based baseline and outputs judged more gender-fair than biased inputs.
Problem
Method
The paper uses real gender-fair text as targets and creates artificial biased sources, including through round-trip translation with biased machine translation models.
Source: §1 · p. 2, §4.2 · p. 5Results
The German model surpasses a strong baseline in WER and its outputs were perceived as more gender-fair than unaltered biased text.
Source: §6 · p. 9Takeaways & Limitations
The results show that biased models can be used to correct bias in gender-fair rewriting without relying on the prevalent handcrafted-rule paradigm.
Source: §6 · pp. 8–9The approach relies on sufficient original gender-fair text in the target language, and the minimum amount of parallel data needed is unknown.
Source: Limitations · p. 9
Abstract
from arXiv · showhide
Natural language generation models reproduce and often amplify the biases present in their training data. Previous research explored using sequence-to-sequence rewriting models to transform biased model outputs (or original texts) into more gender-fair language by creating pseudo training data through linguistic rules. However, this approach is not practical for languages with more complex morphology than English. We hypothesise that creating training data in the reverse direction, i.e. starting from gender-fair text, is easier for morphologically complex languages and show that it matches the performance of state-of-the-art rewriting models for English. To eliminate the rule-based nature of data creation, we instead propose using machine translation models to create gender-biased text from real gender-fair text via round-trip translation. Our approach allows us to train a rewriting model for German without the need for elaborate handcrafted rules. The outputs of this model increased gender-fairness as shown in a human evaluation study.
1 Introduction
The paper addresses persistent gender bias in generated text and the difficulty of creating rule-based rewriting data for morphologically complex languages. It reverses data augmentation and uses biased machine translation to create training data, reporting English benchmark results and a German human-evaluation gain.
1 Introduction
- Generated language models still overproduce masculine forms and perpetuate gender stereotypes, while parallel biased–unbiased text is scarce for training rewriters.Earlier rewriting work addresses the data gap by producing fair targets through handcrafted rules.Source: §1 · pp. 1–2
- Forward Augmentation risks target-side noise and requires rules that are harder to define for languages with complex morphology and grammatical gender.The paper proposes generating artificial biased sources from real gender-fair targets, then using biased off-the-shelf MT to avoid handcrafted rules.Source: Figure 1 · p. 1, §1 · p. 2
- The paper presents reversing augmentation and exploiting biased models as two modifications to rewriting-data creation.The human evaluation reports that the German model's outputs were rated more gender-fair than the original biased texts.Source: §1 · p. 2
- Backward Augmentation matches or outperforms two strong Forward Augmentation baselines in English WER by using human-written unbiased segments as targets.These segments are filtered from large monolingual corpora.Source: §1 · p. 2
- The German rewriter outperforms a heavily engineered rule-based system in WER, and participants rated its outputs as more gender-fair than biased inputs.The human-evaluation campaign involved potential stakeholders.Source: §1 · p. 2
2 Background
Gender-fair rewriting is framed as conditional text generation, but parallel training text is scarce. Rule-based augmentation is relatively tractable in English, while more complex morphology raises challenges; the paper positions round-trip translation as a way to exploit social bias.
2 Background
- Gender-fair rewriting maps biased segments to fair counterparts with sequence-to-sequence models, but such parallel corpora are generally unavailable.Prior work creates artificial fair targets from source text using rule-based augmentation.Source: §2 · p. 2
2.1 Rule-based De-biasing for English
- English rule-based augmentation can use lookup tables, but pronoun changes may require dependency-aware verb changes or resolving context-dependent forms such as “her.”Prior systems address these issues with linguistic tools and language-model scoring, and report WER below 1% on several test sets.Source: §2.1 · pp. 2–3
2.2 Rule-based De-biasing for Other Languages
- For languages with more complex morphology, rule-based pipelines face more frequent rewriting dependencies and harder animacy decisions, and may require extensive linguistic tools.Examples include Spanish and Arabic systems that do not distinguish human referents from objects or require labelled data to identify gendered words.Source: §2.2 · p. 3
2.3 Round-trip Translation
- Earlier round-trip translation applications exploit translation diversity and accuracy or fluency errors; this paper instead exploits machine translation's social biases.The paper describes this as the first use of round-trip translation to exploit social biases.Source: §2.3 · p. 3
3 Backward Augmentation
Backward Augmentation trains a gender-fair rewriter from human-written fair targets and artificially biased sources. Its English evaluation finds WER comparable to Forward Augmentation, with slight gains over prior work on several test sets.
3.1 Method
- The method filters monolingual corpora for gender-fair targets, creates artificial biased sources with reversed rules, and trains a sequence-to-sequence model to predict the fair targets.This reverses the earlier objective of predicting artificial fair targets from original sources.Source: §3.1 · p. 3
3.2 Experimental Setup
- English training data comes from OSCAR, with five million selected segments per augmentation approach and added complementary gendered and non-gendered examples.The data is filtered for duplicates and noise; added examples include masculine-only, feminine-only, and unchanged non-gendered text.Source: §3.2 · p. 3
3.3 Automatic Evaluation
- Evaluation uses published test sets across several domains, tokenised WER, and paired bootstrap resampling for significance testing.The Vanmassenhove et al. test targets were restricted to gender-fair rewriting, excluding other grammatical corrections.Source: §3.3 · pp. 4–5
- Backward Augmentation matches the low WER of the original and the combined Forward Augmentation reproduction, and performs slightly better than prior work on OpenSubtitles, Reddit, and WinoBias+.Table 1 reports tokenised WER, where lower is better.Source: Table 1 · p. 4, §3.3 · p. 5
4 Round-trip Augmentation
Round-trip Augmentation uses machine translation to create biased pseudo-sources from German gender-fair text, avoiding handcrafted de-biasing rules. The combined model achieves the section’s best WER, and human evaluators rate its outputs as more gender-fair than the original texts.
4 Round-trip Augmentation
- Rather than de-biasing source text, Backward Augmentation biases gender-fair targets; this avoids needing animacy prediction for ambiguous German nouns such as “Leiter.”The authors note that rule-based mappings in either direction still require expert knowledge and may not cover all morphologically complex forms.Source: §4 · p. 5
4.1 Method
- The method filters gender-fair German text from monolingual corpora, translates it out and back, then trains a sequence-to-sequence model to map pseudo-sources to fair targets.The training objective is to maximize p(T|Spseudo, θ).Source: §4.1 · p. 5
4.2 Experimental Setup
- Round-trip Augmentation creates pseudo-sources by translating gender-fair German through English, where gender is often lost; about 90% of forms return as generic masculine.The method merges translated counterparts of gender-fair words into the original segment to avoid including unrelated translation differences.Source: §4.2 · p. 5
- Training extensions include merging translation differences, prompting a language model for singular forms, and using gender-controlled translation outputs.With a feminine tag, 36% of produced forms are masculine, compared with 90% from the original checkpoint and 94% with the masculine tag.Source: §4.2 · pp. 5–6
4.3 Automatic Evaluation
- Combining the extensions (+ all) gives the best WER and performs well against Diesner-Mayer and Seidel’s rule-based German rewriter.The authors report that the simpler round-trip model already reduces WER relative to unrevised biased inputs, with further reductions from merging, LM prompting, and gender control.Source: Table 2 · p. 6, §4.3 · p. 6
4.4 Human Evaluation
- The human evaluation used 294 volunteers to rate Original, Rewriter, and manually debiased Gold paragraphs without being told which texts were transformed.In the post-experiment survey, 201 of 294 participants rejected preferring error-free non-gender-fair text; 198 of those rated Rewriter as more gender-fair than Original.Source: §4.4 · p. 7
- Rewriter outputs receive higher mean gender-fairness ratings than Original (2.93 vs. 1.79), though Gold scores higher (3.98).Rewriter also outperforms Original for all six paragraphs and comes close to Gold on P5 and P6.Source: §4.4 · p. 7, Figure 3 · p. 7
5 Discussion
The discussion argues that biased models can help train gender-fair rewriters, while rule-based systems remain limited by their coverage and dependence on language-specific tools. The authors also report that participants preferred potentially erroneous fair wording to error-free biased wording, while noting the German model is imperfect.
5 Discussion
- Biased NLP models can support robust de-biasing rewriters, though the German round-trip experiment covers only one language.The approach assumes biased MT models are available or easier to create than the multiple language-specific tools used in typical rule-based pipelines.Source: §5 · p. 8
- The rule-based baseline handles masculine forms well but performs poorly on feminine forms (WER 22.22), and its rules miss some compounds and unknown gendered words.The authors also identify possible error propagation from the NLP tools used in rule-based approaches.Source: §5 · p. 8
- Rule sets are unlikely to cover every relevant phenomenon, while sequence-to-sequence models offer a model-based alternative for generalization.The authors present round-trip augmentation as a way to create parallel data for new languages without in-depth linguistic knowledge.Source: §5 · p. 8
- Participants rated the German rewriter as more gender-fair than biased originals and said they preferred potentially erroneous fair forms to error-free biased ones.The authors note that the model is far from perfect and argue that such rewriters could be integrated into a wide range of systems with reasonable effort.Source: §5 · p. 8
6 Conclusions and Future Work
The paper uses biased models to train gender-fair rewriters, including a German system whose outputs were rated more gender-fair than the original texts. The authors note broader potential applications where translation models exist.
Conclusions and Future Work
- Biased-model knowledge can be used to correct bias in gender-fair rewriting.The authors also say round-trip translation can substitute for rule-heavy rewriting methods.Source: §6 · pp. 8–9
- In German, the rewriter surpassed a strong baseline in WER, and 294 potential beneficiaries rated its outputs more gender-fair than unaltered biased text.Source: §6 · p. 9
- Where translation models exist, the approach may also apply to dialect rewriting, subjective-bias neutralisation, and discrimination in dialogue systems.Source: §6 · p. 9
Limitations
The approach depends on sufficient original gender-fair text, and the minimum parallel-data requirement is unknown. The survey's scope and use of one German gender-fair form also qualify its findings.
Limitations
- The method requires sufficient original gender-fair text in the language of interest, but the minimum parallel-data amount is unclear.Source: Limitations · p. 9
- The survey targeted affinity groups, limiting generalisability to all German speakers.Source: Limitations · p. 9
- The survey used one form of gender-fair German; other acceptable forms could affect individual ratings, though the authors do not expect them to change the overall finding.Source: Limitations · p. 9
Ethics Statement
Participation was voluntary and anonymous, with no identifying personal data collected. The models were not trained on data filtered to exclude harmful content, though the authors do not expect them to hallucinate it independently.
Ethics Statement
- Participation was voluntary and fully anonymous, and the study collected no personal data that could identify participants.Source: Ethics Statement · p. 9
- Participants were not excluded unless they requested it or reported technical difficulties.Source: Ethics Statement · p. 9
- The rewriting models' publicly available training data was not filtered to remove harmful content; the authors do not believe the models will hallucinate it on their own because they mainly copy text.Source: Ethics Statement · p. 9
A Additional Data Details
The appendix describes filtering and dataset combinations for the English and German rewriters. English variants use rule-based gendered sources, while German variants use round-trip translations from machine-translation checkpoints.
Additional Data Details
- Noisy segments are defined as those that fail the listed OpusFilter filters.Source: A Additional Data Details · p. 13
- Table 4 identifies gender-fair sources as copies of targets before normalisation; English masculine and feminine sources are rule-based, while German sources come from round-trip translations.For German, masculine sources use the original MT checkpoint and feminine sources use the finetuned checkpoint with a feminine tag.Source: Table 4 · p. 13
- English models use either Forward Augmentation with OSCAR masculine, feminine, and gender-fair sources or Backward Augmentation with OSCAR target-side data; English gendered data is capped at 15 million parallel segments.Source: A Additional Data Details · p. 13
- German models include a Round-trip Augmentation combination of OSCAR target-side masculine and gender-fair data; feminine pseudo-sources sometimes duplicate existing segments, yielding fewer additions in gender-control models.Source: A Additional Data Details · p. 13
- The appendix lists additional combinations using LM prompting, gender control, or both, with OSCAR and language-model target-side data.Source: A Additional Data Details · p. 13
- Across dataset combinations, non-gendered OSCAR data makes up 30% of the total parallel data.Source: A Additional Data Details · p. 13
B Detailed Results for German
This appendix details German rewriting results, a merging procedure for round-trip translations, and language-specific resources and patterns. It also shows multilingual language-model prompting examples and outlines the gender-fair forms considered for German.
B Detailed Results for German
- Gender-aware round-trip translation particularly helps feminine inputs, while language-model prompting significantly reduces WER on singular test cases; the rule-based system does not cover feminine forms.The evaluation groups results by generic feminine or masculine input and plural or singular gender-fair output; lower WER is better in Table 5.Source: B Detailed Results for German · p. 13, Table 5 · p. 14
C Merging Algorithm
- The merging algorithm replaces gender-fair words with close matches from round-trip translations, but may introduce grammatical errors on the pseudo-source side.It uses a 0.6 match cutoff and leaves unmatched cases as round-trip translations; source-side errors are not learned as target outputs.Source: C Merging Algorithm · p. 14
D Additional LM Prompting Details
- Multilingual language-model prompts can generate gender-fair forms over long distances, though some forms, such as French “il/elle,” do not include non-binary people.The prompts use a gender-fair determiner and noun following “Ein*e NOUN*in.”Source: D Additional LM Prompting Details · pp. 14–15
E.1 Pronouns
- The English reproduction uses lookup tables in both directions; its pronoun entries map gendered singular pronouns to “they/their/them/theirs/themself.”The tables also cover gender-neutral alternatives for job titles and other nouns.Source: §E.1 · p. 15, E English Lookup Tables · p. 15, §E.2 · p. 15
E.2 Nouns
- The English noun tables list gender-neutral alternatives for job titles and generic “man,” as well as alternatives for unnecessary feminine forms.Examples include “policeman ↔ police officer,” “mankind ↔ humankind,” and “actress ↔ actor.”Source: §E.2 · p. 15
F German Gender-fair Patterns
- German patterns considered include pair, Binnen-I, slash, gap, colon, and star forms; pair, Binnen-I, and slash forms assume binary gender, unlike gap, colon, and star forms.The latter three forms include non-binary people.Source: F German Gender-fair Patterns · pp. 15–16
- Present participles, synonymous neutral nouns, and avoiding gendered words are outside this work’s scope, though the approach could be extended to them.Present participles are gender-neutral only in plural forms, as in “Studierende.”Source: F German Gender-fair Patterns · p. 16
G Gender-Tagged Data With Pair Forms
The authors create gender-tagged German translation data from pair forms and describe its plural and singular construction. The appendix also specifies survey excerpts and how the survey was introduced and shared.
G Gender-Tagged Data With Pair Forms
- Pair forms in parallel data supply feminine and masculine versions of the same noun for gender-aware machine-translation fine-tuning.The authors create contrasting German targets with a simple replacement operation and attach feminine or masculine source tags.Source: G Gender-Tagged Data With Pair Forms · p. 16
- Plural pair forms support contrasting targets because German plural nouns share inflection across genders.The example contrasts “Schülerinnen” with “Schüler.”Source: G Gender-Tagged Data With Pair Forms · p. 16
- For singular pair forms, the authors assign gender labels based on which gendered form appears first because agreement changes are difficult to construct.The approach is not limited to German: pair forms are common in languages including Arabic, French, Russian, and Spanish.Source: G Gender-Tagged Data With Pair Forms · p. 17
H Overview of Survey Texts
- The survey excerpts span six text types, including study regulations, an insurance privacy policy, game instructions, a news article, a company blog, and a medicine insert.The excerpts include generic masculine forms, and one also contains generic feminine forms.Source: H Overview of Survey Texts · p. 17
I List of Survey Contacts
- Survey links were shared through mailing lists and organizations including non-binary and gender-neutral German associations.Listed contacts also include Transgender Network Switzerland and a mailing list.Source: I List of Survey Contacts · p. 17
- The survey invitation described an anonymous 10–15-minute task rating text excerpts for understandability and inclusivity.The invitation was provided in German and English.Source: I List of Survey Contacts · p. 17
J Intended Use
The models are intended for rewriting biased text into possible gender-fair forms, not for identifying a person’s gender or prescribing a particular form. The authors describe them as research-oriented and caution that potential users should test them thoroughly before deployment beyond research.
J Intended Use
- The models rewrite biased text with possible gender-fair forms; they are not intended to identify a person’s gender or prescribe a particular form.Source: J Intended Use · p. 17
- The models are primarily for research, demonstrating that gender-fair rewriting can be trained without language-specific handwritten rules.Source: J Intended Use · p. 19
- Although the survey suggests potential real-world benefits, the models are not error-free and should be thoroughly tested before deployment outside research contexts.Source: J Intended Use · p. 19