Pith. sign in

REVIEW 3 major objections 5 minor 78 references

Explaining GAND: A Resource on Gender-Ambiguous Natural Data & Contrastive Attribution

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read GAND, a new benchmark of 5,047 naturally occurring English sentences that are strictly gender-ambiguous for a singular referent, shows which contextual words steer an MT system's gendered translation and quantifies a masculine default.

desk verdict GAND is a genuinely useful new natural-data benchmark for gender-ambiguous MT, but its core strict-ambiguity guarantee rests on a single annotator's manual pass with no agreement measure — a real, fixable weakness, not a fatal one. read the letter →

arxiv 2607.22546 v1 pith:YCG4FOP5 submitted 2026-05-06 cs.CL

Janiça Hackenbuchner , Jasper Degraeuwe , Arda Tezcan , Joke Daems This is my paper
classification cs.CL
keywords genderbiasmachinetranslationgender-ambiguousdatabenchmarkdatasetcontrastivefeatureattributioninterpretabilitynaturallanguageprocessing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces GAND, a benchmark of 5,047 naturally occurring English sentences that are claimed to be strictly gender-ambiguous for a singular referent, compiled from public web and subtitle corpora through automated filtering and manual vetting. Unlike earlier gender-bias benchmarks built from synthetic templates, GAND provides realistic, linguistically diverse source sentences for studying how machine translation systems assign gender when the source provides no clear cue. On a 1,000-sentence subset translated into German and Spanish, the authors use manually crafted contrastive translations and saliency attribution to show which source words in context push the model toward one gender rather than another. The analysis finds a strong masculine default—masculine translations are on average 50–56% more probable than their feminine counterparts—and that the most influential cues are nouns, verbs, and adjectives syntactically close to the referent, typically within a dependency distance of one or two. If correct, GAND is the first extensive natural-data resource of its kind and a practical tool for diagnosing and mitigating gender bias in translation.

What carries the argument

The central object is the GAND dataset itself—5,047 strictly gender-ambiguous English sentences—combined with a contrastive-translation interpretability pipeline. For each sentence, the model's original translation (target) is paired with a manually constructed translation that flips the referent's gender (foil); the contrastive gradient norm, computed from the difference in next-token probabilities for the two genders, assigns a saliency score to each source token. A threshold of 15% cumulative saliency selects the most influential words, and linguistic analysis uses part-of-speech tags and dependency distances to characterize them.

What would settle it

Independently re-annotate a random sample of the 5,047 GAND sentences with at least two annotators to check whether the referent is strictly gender-ambiguous; if a substantial fraction are judged to contain disambiguating cues—a name, a gender pronoun, or a stereotyped predicate—the core guarantee fails. Alternatively, run the same contrastive-attribution analysis on another open encoder-decoder model: if the masculine default and the local-context saliency pattern do not replicate, the findings are specific to OPUS-MT rather than general to neural MT.

Watch

Extended reading notes

Core claim

The paper's central claim is that GAND is the first large-scale, fully gender-ambiguous natural-language resource for machine translation, with every included sentence referring to a singular entity whose gender is not implied by the source text. Using a randomly selected 1,000-sentence subset translated from English into Spanish and German with the OPUS-MT models, the authors manually create contrastive translations that differ only in the gender of the referent, then compute saliency scores based on the contrastive probability difference. This reveals that the model's choice of masculine versus feminine translation is driven primarily by content words—nouns, verbs, adjectives, and proper n

Load-bearing premise

The load-bearing premise is that every GAND sentence really is strictly gender-ambiguous; this rests on a single annotator's manual vetting of 14,511 candidates, with no second annotator, so a hidden disambiguating cue in any included sentence would corrupt both the benchmark and the attribution findings built on it.

Editorial extensions

If this is right

  • GAND provides a natural-data benchmark for evaluating how different machine translation systems and large language models handle gender ambiguity, moving beyond synthetic templates.
  • The quantified masculine default—roughly 50–56% higher probability for masculine than feminine translations in this setup—gives a concrete baseline for measuring bias and the effect of mitigation strategies.
  • Salient cues are predominantly content words within a dependency distance of one to three from the referent, suggesting that local syntactic context, not distant discourse, drives gender assignment in these models.
  • The contrastive-attribution methodology can flag when a gendered translation is driven by salient stereotyped cues rather than genuine ambiguity, supporting diagnostic rather than normative uses.
  • Preliminary intervention experiments—removing, masking, or flipping salient words—appear to change the translated gender, pointing toward a causal link between the identified cues and model output.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If GAND's strict-ambiguity guarantee holds, it could become a shared testbed for gender-neutral translation strategies across many target languages; however, the single-annotator vetting means independent re-annotation is needed before relying on it as a gold standard.
  • The saliency findings suggest a practical debiasing recipe: editing, masking, or counterfactually flipping high-saliency gendered content words may shift translations away from masculine defaults—this is a testable extension the paper only begins to explore.
  • The roughly half-overlap in salient words between German and Spanish hints that the attribution pattern may be more model-driven than target-language-specific; testing other open encoder-decoder models would clarify whether this is a general property or specific to OPUS-MT.
  • Because the resource is explicitly diagnostic rather than normative, it could support user-facing ambiguity warnings in translation interfaces, alerting readers when a gendered output is not grounded in the source.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces GAND, a dataset of 5047 English sentences claimed to be strictly gender-ambiguous with respect to a singular referent, compiled from C4 and OpenSubtitles using list-based filtering, rule-based cleaning, and manual verification. A 1000-sentence subset (GAND-CT) is translated into German and Spanish with OPUS-MT, manually extended with contrastive translations, and analyzed via gradient-based saliency attribution to identify source words that influence the model's choice of masculine versus feminine target gender. The authors report a masculine-default pattern (higher contrastive probability differences for masculine translations) and find that salient words are predominantly nouns, verbs, and adjectives at dependency distance 1–2 from the referent.

Significance. If the strict gender-ambiguity guarantee holds, GAND fills a genuine gap: existing benchmarks are largely synthetic or include unambiguous sentences, while GAND offers natural, linguistically varied English sentences for studying gender behavior in MT. The contrastive-attribution pipeline is a useful methodological template, and the public release of the dataset and scripts supports reproducibility and future benchmarking. The main risk is that the central ambiguity property is asserted from a single annotator's manual vetting, and the interpretability findings rely on threshold and statistical choices that are not yet fully validated. These concerns are addressable, and the resource itself is a solid contribution if they are resolved.

major comments (3)
  1. [§3.3, Appendix A.2.2] The central property of GAND—strict gender ambiguity for a singular referent—rests entirely on one annotator's manual verification of 14,511 candidates, with no inter-annotator agreement, second pass, or independent audit. The automatic rules in Table 10 target structural coreference with gender pronouns/proper nouns; they cannot remove non-pronominal disambiguating cues such as semantically gendered activities, titles, or world knowledge. Since 65.33% of manually checked candidates were rejected, the manual pass is doing substantial load-bearing work. If a non-negligible fraction of included sentences contain residual disambiguating cues, both the benchmark property and all downstream attribution findings inherit that error. Please report reliability (e.g., Cohen's κ on a sample annotated by a second annotator, or an independent audit with disagreement analysis), and ideally make the pe
  2. [§4.2, §4.4.2] The 15% cumulative-saliency threshold is adopted from Hackenbuchner et al. (2025c) with no sensitivity analysis in this paper. The counts and POS/dependency-distance distributions reported in §4.4.2 are properties of the word set selected by that threshold; a 10% or 20% threshold could change which words are considered salient and therefore affect the main interpretability conclusions. Please report robustness across a range of thresholds, or otherwise demonstrate that the findings are stable with respect to this choice.
  3. [§4.4.1, Table 3] The claim that the model shows higher contrastive probability differences for masculine than feminine translations is supported only by descriptive means. The feminine subsets are small (64 for EN→DE, 65 for EN→ES), and the reported standard deviations (~0.3) overlap substantially with the masculine means. Without confidence intervals, effect sizes, or a significance test, the quantitative comparison is not established. Similarly, the POS and dependency-distance comparisons against the overall distributions in §4.4.2 are not accompanied by any uncertainty quantification. Please add appropriate inferential statistics, or explicitly frame these as descriptive observations for the current sample.
minor comments (5)
  1. [§4.2 heading] Typo: 'Saliency Attibution' should be 'Saliency Attribution'.
  2. [Table 3] The counts for EN→DE (893 + 64 + 41 = 998) and EN→ES (814 + 65 + 120 = 999) do not sum to 1000. Please clarify whether some sentences were excluded for another reason or whether these are rounding artifacts.
  3. [§4.4.1] The sentence beginning 'More frequently than in masculine scenarios, the CPD of an original feminine target was negative...' is difficult to parse. It would be clearer to state explicitly that a negative CPD means the model assigned higher probability to the foil than to the observed target in that contrastive pair.
  4. [Figure 3 caption] The caption references 'red vertical lines' and 'horizontal red lines' without explaining in the caption what the red lines represent; please add this information to the figure caption or legend.
  5. [§3.1, Table 4] The male/female/neutral labels for the 183 referents are derived from word embeddings and an LLM, which encode societal stereotypes. Using these labels to define 'matches' and 'mismatches' in §4.4.1 is descriptively reasonable, but the wording risks reifying the labels as ground truth. Please clarify that these are association-based labels, not objective gender categories.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: GAND's construction is independent of the tested model, and the attribution analysis is a direct measurement; only a minor same-team methodological self-citation appears.

full rationale

The paper's central resource claim—that GAND contains 5047 naturally sourced English sentences that are strictly gender-ambiguous for a singular referent—is not derived from the model being analyzed. Sentences are filtered from C4 and OpenSubtitles using automated rules and manual vetting (Sections 3.1–3.3); the ambiguity property is an annotation decision, not a model output. The interpretability analysis is likewise a direct measurement: OPUS-MT translations are contrasted with manually created gender-contrastive translations, and saliency is computed via the standard contrastive gradient norm of Yin and Neubig (2022). The finding that nouns, verbs, and adjectives near the referent are salient is an empirical result of that computation, not an input to it. The only same-team self-citation that could be flagged is the 15% cumulative-saliency threshold and the preprocessing choices taken from Hackenbuchner et al. (2025c). This is a methodological hyperparameter inherited from prior work, explicitly described as replaceable ('Future work could include analyses based on alternative thresholds'), and the main conclusions do not reduce to that threshold. A substantive correctness risk, but not a circularity, is the single-annotator manual verification of gender ambiguity without inter-annotator agreement; if label noise exists, downstream attribution findings inherit it, but this is a validity concern rather than a derivation that assumes its own conclusion. Overall, the paper contains no step where a 'prediction' is equivalent by construction to a fitted input or where the central result is forced by a self-citation chain.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The central claims rest on the hand-curated referent lists, manual vetting, and interpretability assumptions listed above. None are machine-checked in this paper.

free parameters (5)
  • 183-referent gender-association list = 72 male-associated, 57 female-associated, 54 neutral
    Section 3.1: The set of referent nouns and their binary/neutral gender labels, compiled from prior embedding studies and ChatGPT-4, determines which sentences can enter GAND and drives the match/mismatch analysis.
  • Sentence length cutoffs = min 5 tokens, max 50 tokens
    Section 3.1: Hand-chosen filtering thresholds that shape the dataset's sentence distribution.
  • Coreference exclusion rules = 10 handcrafted rules plus manually compiled gender (pro)noun list
    Section 3.2 / Appendix Table 10: These rules encode what counts as a disambiguating cue; different rules would yield a different dataset.
  • Saliency threshold = 15% cumulative attribution (top-15% subset)
    Section 4.2: The minimum subset of source words whose cumulative attribution reaches 15% of total defines 'salient'; threshold is imported from Hackenbuchner et al. (2025c) and no sensitivity analysis is given.
  • Token removal set for attribution = target referent token, EOS, punctuation, and {a, an, the, this, that, these, those}
    Section 4.2: Preprocessing choice that removes the referent itself and function words before aggregating saliency; affects which words are reported as salient.
assumptions (6)
  • domain assumption English sentences containing one of 183 hand-selected referent nouns, after automatic filtering and single-annotator manual vetting, are strictly gender-ambiguous with respect to that referent.
    Section 3.3: manual verification by the main author is the final arbiter of ambiguity; no inter-annotator agreement or external validation is reported, so the dataset's core property is assumed from one annotator's judgment.
  • domain assumption The gender association labels for referent nouns (masculine/feminine/neutral) derived from prior word-embedding studies and ChatGPT-4 reflect meaningful genderedness of the referent.
    Section 3.1: referents such as 'assistant' (female) and 'user' (male) are taken from Bolukbasi et al. 2016 / Stanovsky et al. 2019 / Caliskan et al. 2022 and LLM neutrality checks; these labels drive match/mismatch analyses in Section 4.4.1.
  • domain assumption Saliency attribution (gradient-norm contrastive explanations from inseq) reflects the contextual cues that actually inform the model's gender choice.
    Section 4.2: the analysis equates high attribution scores with influence; the paper itself treats causal intervention as future work, so attribution is assumed to be explanatory.
  • domain assumption Manually created contrastive translations differ from the original only in gender-marked target tokens, so the attribution contrast isolates gender.
    Section 4.1: authors note that 'more than one term must frequently be flipped' to contrast gender, but the analysis left-aligns at the first differing token; semantic invariance is assumed.
  • ad hoc to paper The 15% cumulative-saliency threshold yields plausible salient cues.
    Section 4.2: threshold adopted from Hackenbuchner et al. (2025c) rather than re-derived on GAND; no sensitivity analysis.
  • domain assumption Stanza and spaCy dependency parses are accurate enough for coreference exclusion checks and dependency-distance measurements.
    Sections 3.2, 4.3: automated cleaning and the distance analysis rely on parser outputs; no parser error analysis is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Explaining GAND: A Resource on Gender-Ambiguous Natural Data & Contrastive Attribution." pith.science (2026). https://pith.science/paper/YCG4FOP5

@misc{pith2026260722546,
  author       = {Pith},
  title        = {Pith review of: Explaining GAND: A Resource on Gender-Ambiguous Natural Data & Contrastive Attribution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YCG4FOP5}},
  note         = {Machine review of arXiv:2607.22546}
}
read the original abstract

Machine translation (MT) systems continue to produce gender-biased translations. In a time where self-expression is paramount, mistranslations based on default behaviour and stereotyping can lead to harm for users of these systems. To better understand how these systems translate gender in the absence of clear gender cues, we need benchmarking resources that reflect gender-ambiguous scenarios in a natural way. To this end, we present GAND, a gender-ambiguous natural data benchmarking resource for MT consisting of English source sentences, specifically designed to analyse the influence of contextual cues on gender in translation. We leverage GAND to conduct an interpretability analysis: we translate a subset of GAND into two grammatical gender languages and extend these with manually crafted contrastive translations. A following feature attribution analysis reveals source words in context that inform the gender translation of an ambiguous referent entity in the target translation.

Figures

Figures reproduced from arXiv: 2607.22546 by the authors.

Figure 1
Figure 1. Infographic of the creation of GAND and a simpli￾fied visualisation of the interpretability analysis via contrastive translations. (e.g., English) to grammatical languages with overt gender distinctions (e.g., German or Spanish), MT systems tend to assign grammatical gender to am￾biguous referent entities. This is a phenomenon that Rarrick et al. (2023, p. 845) dub “Arbitrarily Gender-Marked Entities”, where referen… view at source ↗
Figure 2
Figure 2. Contrastive probability difference of the model predicting one (gendered) token instead of the other. differences when translating a referent as mascu￾line, with a mean of 0.56 for EN-DE, meaning that the masculine translation of a referent entity is 56% more probable than the feminine one, and on aver￾age 50% more probable for EN-ES. In comparison, the model shows, on average, lower CPDs when translating a referent… view at source ↗
Figure 3
Figure 3. Parts-of-speech distribution and dependency dis￾tances of salient words. ∼8% (ES 6.7%; DE 8.8%), and pronouns with ∼7% (ES 6.7%; DE 7.3%). This is visualised in [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Mean probability difference per referent for Spanish, where there is a mismatch in referent embedding and target gender (e.g., neutral ‘embedding’ but masculine target translation) [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Mean probability difference per referent for German, where there is a mismatch in referent embedding and target gender (e.g., neutral ‘embedding’ but masculine target translation) [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Mean probability difference per referent for Spanish, where there is a match in referent embedding and target gender (e.g., feminine embedding and feminine target translation) [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Mean probability difference per referent for German, where there is a match in referent embedding and target gender (e.g., feminine embedding and feminine target translation) [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Heatmap of POS tags of salient words for each referent for DE and ES. The x-axis shows POS categories, the y-axis shows (source) referents. Referents that have been translated into masculine are depicted in blue, into feminine are depicted in orange, and into neutral i…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

78 extracted references · 21 canonical work pages

  1. [1]

    A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation

    Attanasio, Giuseppe and Plaza del Arco, Flor Miriam and Nozza, Debora and Lauscher, Anne. A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.243

    doi:10.18653/v1/2023.emnlp-main.243 2023

  2. [2]

    Hi Guys or Hi Folks? Benchmarking Gender-Neutral Machine Translation with the G e NTE Corpus

    Piergentili, Andrea and Savoldi, Beatrice and Fucci, Dennis and Negri, Matteo and Bentivogli, Luisa. Hi Guys or Hi Folks? Benchmarking Gender-Neutral Machine Translation with the G e NTE Corpus. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.873

    work page doi:10.18653/v1/2023.emnlp-main.873 2023

  3. [3]

    Interpreting Language Models with Contrastive Explanations

    Yin, Kayo and Neubig, Graham. Interpreting Language Models with Contrastive Explanations. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.14

    doi:10.18653/v1/2022.emnlp-main.14 2022

  4. [4]

    The elephant in the interpretability room:

    Bastings, Jasmijn and Filippova, Katja , year =. The elephant in the interpretability room:. Proceedings of the. doi:10.18653/v1/2020.blackboxnlp-1.14 , language =

    doi:10.18653/v1/2020.blackboxnlp-1.14 2020

  5. [5]

    Yin, Kayo and Fernandes, Patrick and Pruthi, Danish and Chaudhary, Aditi and Martins, André F. T. and Neubig, Graham , year =. Do. Proceedings of the 59th. doi:10.18653/v1/2021.acl-long.65 , language =

    doi:10.18653/v1/2021.acl-long.65 2021

  6. [6]

    Bastings, Jasmijn and Ebert, Sebastian and Zablotskaia, Polina and Sandholm, Anders and Filippova, Katja , year =. “. Proceedings of the 2022. doi:10.18653/v1/2022.emnlp-main.64 , language =

    work page doi:10.18653/v1/2022.emnlp-main.64 2022

  7. [7]

    Are We Paying Attention to Her? Investigating Gender Disambiguation and Attention in Machine Translation

    Manna, Chiara and Alishahi, Afra and Blain, Fr \'e d \'e ric and Vanmassenhove, Eva. Are We Paying Attention to Her? Investigating Gender Disambiguation and Attention in Machine Translation. Proceedings of the 3rd Workshop on Gender-Inclusive Translation Technologies (GITT 2025). 2025

    2025

  8. [8]

    Inseq: An Interpretability Toolkit for Sequence Generation Models

    Sarti, Gabriele and Feldhus, Nils and Sickert, Ludwig and van der Wal, Oskar. Inseq: An Interpretability Toolkit for Sequence Generation Models. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations). 2023. doi:10.18653/v1/2023.acl-demo.40

    doi:10.18653/v1/2023.acl-demo.40 2023

Show all 78 references
  1. [9]

    A decade of gender bias in machine translation

    Savoldi, Beatrice and Bastings, Jasmijn and Bentivogli, Luisa and Vanmassenhove, Eva. A decade of gender bias in machine translation. Patterns. 2025. doi:10.1016/j.patter.2025.101257

    2025

  2. [10]

    2024 , eprint=

    A Primer on the Inner Workings of Transformer-based Language Models , author=. 2024 , eprint=

    2024

  3. [11]

    Getting Gender Right in Neural Machine Translation

    Vanmassenhove, Eva and Hardmeier, Christian and Way, Andy. Getting Gender Right in Neural Machine Translation. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing. 2018. doi:10.18653/v1/D18-1334

    2018 doi

  4. [12]

    Analysing Neural Language Models: Contextual Decomposition Reveals Default Reasoning in Number and Gender Assignment

    Jumelet, Jaap and Zuidema, Willem and Hupkes, Dieuwke. Analysing Neural Language Models: Contextual Decomposition Reveals Default Reasoning in Number and Gender Assignment. Proceedings of the 23rd Conference on Computational Natural Language Learning (CoNLL). 2019. doi:10.1865...

    2019 doi

  5. [13]

    First the Worst: Finding Better Gender Translations During Beam Search

    Saunders, Danielle and Sallis, Rosie and Byrne, Bill. First the Worst: Finding Better Gender Translations During Beam Search. Findings of the Association for Computational Linguistics: ACL 2022. 2022. doi:10.18653/v1/2022.findings-acl.301

    2022 doi

  6. [14]

    2023 , eprint=

    Toward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networks , author=. 2023 , eprint=

    2023

  7. [15]

    The Unheard Alternative: Contrastive Explanations for Speech-to-Text Models

    Conti, Lina and Fucci, Dennis and Gaido, Marco and Negri, Matteo and Wisniewski, Guillaume and Bentivogli, Luisa. The Unheard Alternative: Contrastive Explanations for Speech-to-Text Models. Proceedings of the 8th BlackboxNLP Workshop: Analyzing and Interpreting Neural Network...

    2025 doi

  8. [16]

    https://aclanthology.org/2020.eamt-1.61/

    J. Proceedings of the 22nd Annual Conferenec of the European Association for Machine Translation (EAMT) , url="https://aclanthology.org/2020.eamt-1.61/", year =

    2020

  9. [17]

    Gender Bias and the Role of Context in Human Perception and Machine Translation

    Hackenbuchner, Jani c a and Tezcan, Arda and Daems, Joke. Gender Bias and the Role of Context in Human Perception and Machine Translation. Computational Linguistics in the Netherlands Journal. 2025

    2025

  10. [18]

    Automatic detection of (potential) factors in the source text leading to gender bias in machine translation

    Hackenbuchner, Jani c a and Tezcan, Arda and Daems, Joke. Automatic detection of (potential) factors in the source text leading to gender bias in machine translation. Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 2). 2024

    2024

  11. [19]

    James Murdoch and Chandan Singh and Karl Kumbier and Reza Abbasi-Asl and Bin Yu

    W. James Murdoch and Chandan Singh and Karl Kumbier and Reza Abbasi-Asl and Bin Yu. Definitions, methods, and applications in interpretable machine learning. Proceedings of the National Academy of Sciences. 2019. doi:10.1073/pnas.1900654116. https://www.pnas.org/doi/pdf/10.107...

    2019 doi

  12. [20]

    Contrastive Explanation

    Lipton, Peter. Contrastive Explanation. Royal Institute of Philosophy Supplement. 1990. doi:10.1017/S1358246100005130

    1990 doi

  13. [21]

    and Garcke, Jochen , journal=

    Roscher, Ribana and Bohn, Bastian and Duarte, Marco F. and Garcke, Jochen , journal=. Explainable Machine Learning for Scientific Insights and Discoveries , year=

  14. [22]

    Biran, Or, and Courtenay V. Cotton. Explanation and Justification in Machine Learning: A Survey. Proceedings of the IJCAI-17 Workshop on Explainable Artificial Intelligence (XAI). 2017

    2017

  15. [23]

    Interpretable Machine Learning , subtitle=

    Christoph Molnar , year=. Interpretable Machine Learning , subtitle=

  16. [24]

    Proceedings of The Fourth Widening Natural Language Processing Workshop , publisher=

    Towards Mitigating Gender Bias in a decoder-based Neural Machine Translation model by Adding Contextual Information , author=. Proceedings of The Fourth Widening Natural Language Processing Workshop , publisher=

  17. [25]

    , month = jul, year =

    Caliskan, Aylin and Ajay, Pimparkar Parth and Charlesworth, Tessa and Wolfe, Robert and Banaji, Mahzarin R. , month = jul, year =. Gender. Proceedings of the 2022. doi:10.1145/3514094.3534162 , language =

    2022

  18. [26]

    https://dl.acm.org/doi/10.5555/3157382.3157584

    Man is to. CoRR , url="https://dl.acm.org/doi/10.5555/3157382.3157584", author =

  19. [27]

    Zhao, Jieyu and Wang, Tianlu and Yatskar, Mark and Cotterell, Ryan and Ordonez, Vicente and Chang, Kai-Wei , month = apr, year =. Gender. doi:10.48550/arXiv.1904.03310 , language =

    1904 doi arXiv

  20. [28]

    Proceedings of the 5th Conference on Machine Translation (WMT) , publisher=

    Gender Coreference and Bias Evaluation at WMT 2020 , author=. Proceedings of the 5th Conference on Machine Translation (WMT) , publisher=

    2020

  21. [29]

    Collecting a

    Levy, Shahar and Lazar, Koren and Stanovsky, Gabriel , year =. Collecting a. Findings of the. doi:10.18653/v1/2021.findings-emnlp.211 , language =

    2021 doi

  22. [30]

    Gendered Technology in Translation and Interpreting , pages=

    Gender bias in machine translation and the era of large language models , author=. Gendered Technology in Translation and Interpreting , pages=. 2024 , publisher=

    2024

  23. [31]

    g EN der- IT : An Annotated E nglish- I talian Parallel Challenge Set for Cross-Linguistic Natural Gender Phenomena

    Vanmassenhove, Eva and Monti, Johanna. g EN der- IT : An Annotated E nglish- I talian Parallel Challenge Set for Cross-Linguistic Natural Gender Phenomena. Proceedings of the 3rd Workshop on Gender Bias in Natural Language Processing. 2021. doi:10.18653/v1/2021.gebnlp-1.1

    2021 doi

  24. [32]

    Mind the Inclusivity Gap: Multilingual Gender-Neutral Translation Evaluation with m G e NTE

    Savoldi, Beatrice and Attanasio, Giuseppe and Cupin, Eleonora and Gkovedarou, Eleni and Hackenbuchner, Jani c a and Lauscher, Anne and Negri, Matteo and Piergentili, Andrea and Thind, Manjinder and Bentivogli, Luisa. Mind the Inclusivity Gap: Multilingual Gender-Neutral Transl...

    2025 doi

  25. [33]

    2025 , eprint=

    What Triggers my Model? Contrastive Explanations Inform Gender Choices by Translation Models , author=. 2025 , eprint=

    2025

  26. [34]

    Gender bias amplification during Speed-Quality optimization in Neural Machine Translation

    Renduchintala, Adithya and Diaz, Denise and Heafield, Kenneth and Li, Xian and Diab, Mona. Gender bias amplification during Speed-Quality optimization in Neural Machine Translation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the...

    2021 doi

  27. [35]

    M i TT en S : A Dataset for Evaluating Gender Mistranslation

    Robinson, Kevin and Kudugunta, Sneha and Stella, Romina and Dev, Sunipa and Bastings, Jasmijn. M i TT en S : A Dataset for Evaluating Gender Mistranslation. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp...

    2024 doi

  28. [36]

    Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods

    Zhao, Jieyu and Wang, Tianlu and Yatskar, Mark and Ordonez, Vicente and Chang, Kai-Wei. Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods. Proceedings of the 2018 Conference of the North A merican Chapter of the Association for Computational Linguistics: ...

    2018 doi

  29. [37]

    Automatic Gender Identification and Reinflection in A rabic

    Habash, Nizar and Bouamor, Houda and Chung, Christine. Automatic Gender Identification and Reinflection in A rabic. Proceedings of the First Workshop on Gender Bias in Natural Language Processing. 2019. doi:10.18653/v1/W19-3822

    2019 doi

  30. [38]

    2023 , isbn =

    Rarrick, Spencer and Naik, Ranjita and Mathur, Varun and Poudel, Sundar and Chowdhary, Vishal , title =. 2023 , isbn =. doi:10.1145/3600211.3604675 , booktitle =

    2023

  31. [39]

    Analyzing Gender Translation Errors to Identify Information Flows between the Encoder and Decoder of a NMT System

    Wisniewski, Guillaume and Zhu, Lichao and Ballier, Nicolas and Yvon, Fran c ois. Analyzing Gender Translation Errors to Identify Information Flows between the Encoder and Decoder of a NMT System. Proceedings of the Fifth BlackboxNLP Workshop on Analyzing and Interpreting Neura...

    2022 doi

  32. [40]

    Assumed Identities: Quantifying Gender Bias in Machine Translation of Gender-Ambiguous Occupational Terms

    Mastromichalakis, Orfeas Menis and Filandrianos, Giorgos and Symeonaki, Maria and Stamou, Giorgos. Assumed Identities: Quantifying Gender Bias in Machine Translation of Gender-Ambiguous Occupational Terms. Proceedings of the 2025 Conference on Empirical Methods in Natural Lang...

    2025 doi

  33. [41]

    2025 , eprint=

    Voice, Bias, and Coreference: An Interpretability Study of Gender in Speech Translation , author=. 2025 , eprint=

    2025

  34. [42]

    Gender Bias in Machine Translation

    Savoldi, Beatrice and Gaido, Marco and Bentivogli, Luisa and Negri, Matteo and Turchi, Marco. Gender Bias in Machine Translation. Transactions of the Association for Computational Linguistics. 2021. doi:10.1162/tacl_a_00401

    2021 doi

  35. [43]

    Language (

    Blodgett, Su Lin and Barocas, Solon and Daumé III, Hal and Wallach, Hanna , editor =. Language (. Proceedings of the 58th. 2020 , pages =. doi:10.18653/v1/2020.acl-main.485 , urldate =

    2020 doi

  36. [44]

    Context-Aware Neural Machine Translation Learns Anaphora Resolution

    Voita, Elena and Serdyukov, Pavel and Sennrich, Rico and Titov, Ivan. Context-Aware Neural Machine Translation Learns Anaphora Resolution. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2018. doi:10.18653/v1/P18-1117

    2018 doi

  37. [45]

    Quantifying the

    Sarti, Gabriele and Chrupał a, Grzegorz and Nissim, Malvina and Bisazza, Arianna , editor =. Quantifying the. International. 2024 , pages =

    2024

  38. [46]

    Sennrich, Rico , year =. How. Proceedings of the 15th. doi:10.18653/v1/E17-2060 , language =

    doi

  39. [47]

    Improving

    Rios Gonzales, Annette and Mascarell, Laura and Sennrich, Rico , year =. Improving. Proceedings of the. doi:10.18653/v1/W17-4702 , language =

    doi

  40. [48]

    , author=

    Measuring nominal scale agreement among many raters. , author=. Psychological bulletin , volume=. 1971 , publisher=

    1971

  41. [49]

    Educational and psychological measurement , volume=

    A coefficient of agreement for nominal scales , author=. Educational and psychological measurement , volume=. 1960 , publisher=

    1960

  42. [50]

    You Shall Know a Word ' s Gender by the Company it Keeps: Comparing the Role of Context in Human Gender Assumptions with MT

    Hackenbuchner, Jani c a and Daems, Joke and Tezcan, Arda and Maladry, Aaron. You Shall Know a Word ' s Gender by the Company it Keeps: Comparing the Role of Context in Human Gender Assumptions with MT. Proceedings of the 2nd International Workshop on Gender-Inclusive Translati...

    2024

  43. [51]

    Proceedings of the AAAI Conference on Artificial Intelligence , author =

    Interpreting. Proceedings of the AAAI Conference on Artificial Intelligence , author =. 2022 , pages =. doi:10.1609/aaai.v36i11.21442 , language =

    2022 doi

  44. [52]

    Lage, Isaac and Chen, Emily and He, Jeffrey and Narayanan, Menaka and Kim, Been and Gershman, Sam and Doshi-Velez, Finale , month = aug, year =. An. doi:10.48550/arXiv.1902.00006 , language =

    1902 doi arXiv

  45. [53]

    Contrastive

    Vamvas, Jannis and Sennrich, Rico , year =. Contrastive. Proceedings of the 2021. doi:10.18653/v1/2021.emnlp-main.803 , language =

    2021 doi

  46. [54]

    Jacovi, Alon and Goldberg, Yoav , year =. Towards. Proceedings of the 58th. doi:10.18653/v1/2020.acl-main.386 , language =

    2020 doi

  47. [55]

    ACM Computing Surveys , author =

    Post-hoc. ACM Computing Surveys , author =. 2022 , pages =. doi:10.1145/3546577 , language =

    2022 doi

  48. [56]

    Implementing Gender-Inclusivity in MT Output using Automatic Post-Editing with LLM s

    Nunziatini, Mara and Diego, Sara. Implementing Gender-Inclusivity in MT Output using Automatic Post-Editing with LLM s. Proceedings of the 25th Annual Conference of the European Association for Machine Translation (Volume 1). 2024

    2024

  49. [57]

    N eu T ral R ewriter: A Rule-Based and Neural Approach to Automatic Rewriting into Gender Neutral Alternatives

    Vanmassenhove, Eva and Emmery, Chris and Shterionov, Dimitar. N eu T ral R ewriter: A Rule-Based and Neural Approach to Automatic Rewriting into Gender Neutral Alternatives. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. doi:10.18...

    2021 doi

  50. [58]

    A Rewriting Approach for Gender Inclusivity in P ortuguese

    Veloso, Leonor and Coheur, Luisa and Ribeiro, Rui. A Rewriting Approach for Gender Inclusivity in P ortuguese. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023. doi:10.18653/v1/2023.findings-emnlp.585

    2023 doi

  51. [59]

    Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model

    Amrhein, Chantal and Schottmann, Florian and Sennrich, Rico and L. Exploiting Biased Models to De-bias Text: A Gender-Fair Rewriting Model. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023. doi:10.18653/v1/20...

    2023 doi

  52. [60]

    Gender, names and other mysteries: Towards the ambiguous for gender-inclusive translation

    Saunders, Danielle and Olsen, Katrina. Gender, names and other mysteries: Towards the ambiguous for gender-inclusive translation. Proceedings of the First Workshop on Gender-Inclusive Translation Technologies. 2023

    2023

  53. [61]

    Cho, Won Ik and Kim, Ji Won and Kim, Seok Min and Kim, Nam Soo , year =. On. Proceedings of the. doi:10.18653/v1/W19-3824 , language =

    doi

  54. [62]

    Multilingual Holistic Bias: Extending Descriptors and Patterns to Unveil Demographic Biases in Languages at Scale

    Costa-juss \`a , Marta and Andrews, Pierre and Smith, Eric and Hansanti, Prangthip and Ropers, Christophe and Kalbassi, Elahe and Gao, Cynthia and Licht, Daniel and Wood, Carleigh. Multilingual Holistic Bias: Extending Descriptors and Patterns to Unveil Demographic Biases in L...

    2023 doi

  55. [63]

    and Cattoni, Roldano and Turchi, Marco

    Bentivogli, Luisa and Savoldi, Beatrice and Negri, Matteo and Di Gangi, Mattia A. and Cattoni, Roldano and Turchi, Marco. Gender in Danger? Evaluating Speech Translation Technology on the M u ST - SHE Corpus. Proceedings of the 58th Annual Meeting of the Association for Comput...

    2020 doi

  56. [64]

    Reducing

    Saunders, Danielle and Byrne, Bill , year =. Reducing. Proceedings of the 58th. doi:10.18653/v1/2020.acl-main.690 , language =

    2020 doi

  57. [65]

    Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics , volume=

    Savoldi, Beatrice and Piergentili, Aandrea and Fucci, Dennis and Negri, Matteo and Bentivogli, Luisa , title=. Proceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics , volume=

  58. [66]

    GAMBIT +: A Challenge Set for Evaluating Gender Bias in Machine Translation Quality Estimation Metrics

    Filandrianos, George and Menis Mastromichalakis, Orfeas and Mohammed, Wafaa and Attanasio, Giuseppe and Zerva, Chrysoula. GAMBIT +: A Challenge Set for Evaluating Gender Bias in Machine Translation Quality Estimation Metrics. Proceedings of the Tenth Conference on Machine Tran...

    2025 doi

  59. [67]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing) , publisher=

    Prequel: Quality estimation of machine translation outputs in advance , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing) , publisher=

    2022

  60. [68]

    Proceedings of the Fourth Workshop on Discourse in Machine Translation (DiscoMT) , publisher=

    When and Why is Document-level Context Useful in Neural Machine Translation? , author=. Proceedings of the Fourth Workshop on Discourse in Machine Translation (DiscoMT) , publisher=

  61. [69]

    and Zettlemoyer, Luke , year =

    Stanovsky, Gabriel and Smith, Noah A. and Zettlemoyer, Luke , year =. Evaluating. Proceedings of the 57th. doi:10.18653/v1/P19-1164 , language =

    doi

  62. [70]

    Gender bias and stereotypes in

    Kotek, Hadas and Dockum, Rikker and Sun, David , month = nov, year =. Gender bias and stereotypes in. Proceedings of. doi:10.1145/3582269.3615599 , language =

  63. [71]

    and Casas, Noe , year =

    Basta, Christine and Costa-jussà, Marta R. and Casas, Noe , year =. Evaluating the. Proceedings of the. doi:10.18653/v1/W19-3805 , language =

    doi

  64. [72]

    MT - G en E val: A Counterfactual and Contextual Dataset for Evaluating Gender Accuracy in Machine Translation

    Currey, Anna and Nadejde, Maria and Pappagari, Raghavendra Reddy and Mayer, Mia and Lauly, Stanislas and Niu, Xing and Hsu, Benjamin and Dinu, Georgiana. MT - G en E val: A Counterfactual and Contextual Dataset for Evaluating Gender Accuracy in Machine Translation. Proceedings...

    2022 doi

  65. [73]

    Sewunetie, Walelign Tewabe and Tonja, Atnafu Lambebo and Belay, Tadesse Destaw and Nigatu, Hellina Hailu and Kidanu, Gashaw and Mossie, Zewdie and Seid, Hussien and Yimam, Seid Muhie , year=. Gender. Proceedings of the 2nd International Workshop on Gender-Inclusive Translation...

  66. [74]

    Gkovedarou, Eleni and Daems, Joke and Bruyne, Luna De , year =. Gender. Proceedings of the 3rd Workshop on Gender-Inclusive Translation Technologies (GITT 2025) , publisher =

    2025

  67. [75]

    Gender Neutralization for an Inclusive Machine Translation: from Theoretical Foundations to Open Challenges

    Piergentili, Andrea and Fucci, Dennis and Savoldi, Beatrice and Bentivogli, Luisa and Negri, Matteo. Gender Neutralization for an Inclusive Machine Translation: from Theoretical Foundations to Open Challenges. Proceedings of the First Workshop on Gender-Inclusive Translation T...

    2023

  68. [76]

    Perspectives , author =

    Gender-fair translation: a case study beyond the binary , volume =. Perspectives , author =. 2024 , pages =. doi:10.1080/0907676X.2023.2268654 , language =

    2024

  69. [77]

    Glitter: A Multi-Sentence, Multi-Reference Benchmark for Gender-Fair G erman Machine Translation

    Pranav, A and Hackenbuchner, Jani c a and Attanasio, Giuseppe and Lardelli, Manuel and Lauscher, Anne. Glitter: A Multi-Sentence, Multi-Reference Benchmark for Gender-Fair G erman Machine Translation. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025....

    2025 doi

  70. [78]

    , year =

    Qi, Peng and Zhang, Yuhao and Zhang, Yuhui and Bolton, Jason and Manning, Christopher D. , year =. Stanza:. doi:10.48550/ARXIV.2003.07082 , abstract =

    2003 doi arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.