{"id":"38f660b6-9774-4d4b-856c-b65fa1fee418","arxiv_id":"2411.12074","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Regularized masked-language modeling and name-masking reduce gender bias in embeddings, but the contextual results rely heavily on evaluation metrics aligned with the training objective.","lead":"The authors propose regularized masked-language-model and name-masking training recipes to reduce gender bias in BERT and word2vec-style embeddings, reporting large drops on their own bias probes. The paper also argues, with supporting experiments, that stereotypical names rather than gendered words are the main carriers of gender bias in static embeddings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contextual debiasing only blocks attention to attribute words during training; no probe tests whether gender is recoverable from the remaining context at inference, so the central claim may be superficial.","rationale":"The reader's weakest assumption correctly identifies the soft spot: if gender can be recovered from the remaining context, the contextual debiasing is superficial in exactly the way the paper criticizes for static embeddings. This is more fundamental than the missing training details or the circularity of the intrinsic metric, because it attacks the causal mechanism claimed for the method. The regularizer directly optimizes the same prediction scores used in the main evaluation, so the improvement in Figs. 2 and 3 may reflect optimization of the metric rather than a genuine removal of gender information from the representations. The absence of any probing or recovery experiment leaves this as a live possibility, and the paper's own static analyses showing that names still carry gender after debiasing strengthen the concern. I agree with the reader's REJECT: the contextual contribution is not adequately supported, and the static claim, while interesting, is overclaimed as 'empirical proof' given the small gains and the lack of control for general token masking. No adjustment to the verdict is needed; the existing reject stands.","tokens_in":15154,"tokens_out":4865,"duration_ms":50864,"concrete_test":"Run a gender-recovery probe on the debiased BERT. Take Winogender-style templates in which matched sentences differ only by the pronoun (he/she), and extract the contextual embedding of the occupation token, or of the [MASK] position with the pronoun left unmasked, from the final model. Train a logistic regression to predict whether the original context used he or she from that single embedding. If held-out accuracy is clearly above chance (e.g., above 60%, with confidence intervals), gender information is recoverable from non-attribute context and the debiasing is at least partially superficial. Compare against the original BERT baseline and repeat on templates containing stereotypical names.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that the regularized MLM objective (Section 3.2.1) eliminates the dependencies between attribute words and target words in contextual embeddings. This is not established. The mechanism in Fig. 1 is a training-time attention mask: while predicting a masked noun, the model cannot attend to attribute words, and while predicting a masked attribute word, it cannot attend to nouns. But at evaluation, and in any downstream use, the model is unmasked. Nothing in the objective prevents gender information from being routed through the remaining tokens, such as names, other nouns, adjectives, or verbs, so the target noun's contextual representation can still be gendered. The regularizer in Eq. (1) equalizes prediction scores f(x_i) and f(y_i) under the masked training distribution, and the evaluation in Section 3.4.1 measures exactly those prediction scores (Figs. 2 and 3); this is closer to checking the training objective than to testing the deployed model. This is the same recoverability failure the paper invokes for static embeddings in Section 4.1 via Gonen and Goldberg, but no probing or recovery experiment is reported for BERT. The static experiments themselves show that names remain strong gender carriers even after debiasing (Tables 7 and 8), so attribute words are demonstrably not the only conduit.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes two families of debiasing methods. For contextual embeddings (BERT), it introduces a regularized masked-language-model objective (Eq. 1) with attention masks that prevent the model from using attribute words when predicting nouns and vice versa, plus a gender-prediction data-augmentation strategy. It evaluates these with pronoun/occupation prediction tasks, CrowS-pairs-style scoring, SEAT, and downstream GLUE tasks. For static embeddings, it proposes gender-neutral corpus rewriting, explicit gender encoding (EGE), and masking stereotypical names (NER-M); the headline static result is that NER-M plus EGE reduces Gonen-Goldberg clustering accuracy from about 0.58 to 0.5157. The paper also argues that stereotypical names, not gendered words, are the main source of static gender bias.","tokens_in":15317,"tokens_out":7389,"duration_ms":70857,"significance":"If correct, the regularized MLM would be a practical training-time debiaser for BERT, and the NER-M+EGE result would provide a striking confirmation of the names-as-carriers hypothesis. The paper is honest about some limitations and makes a falsifiable prediction in the clustering experiment. However, the contextual claim is not yet supported because the main intrinsic metric is essentially the training objective and no recovery or probing experiment is reported; the static causal claim is undercut by the ablation pattern. With additional controls and targeted experiments, the ideas could become a contribution, but in the present form the central claims outrun the evidence.","major_comments":[{"comment":"The first intrinsic evaluation is circular with respect to the training objective. Eq. (1) regularizes the model so that f(x_i) and f(y_i) are close for paired gender tokens, and Fig. 2 reports exactly these prediction scores for he/she, him/her, his/her across occupations. The attention mask in Fig. 1 is applied only during continued pre-training; at inference the model is unmasked, so gender information can be rerouted through names, other nouns, adjectives, or verbs. The paper itself invokes the Gonen-Goldberg recoverability critique for static embeddings in Section 4.1 and shows in Tables 7 and 8 that names remain strong gender carriers, yet no probing or recovery experiment is run on BERT. Without a probe (e.g., training a gender classifier on the contextual representations of a masked occupation from the Winogender templates), the abstract's claim that the objective eliminates the dependencies is unsupported.","section":"Section 3.2.1, Eq. (1), Fig. 1; Section 3.4.1"},{"comment":"The static claim that names are the main source of bias is not supported by the ablations. NER masking alone yields 0.5839 ± 0.0356, no better than the original 0.5802 ± 0.0282; EGE alone yields 0.5743 ± 0.0106; only the combination NER-M+EGE drops to 0.5157 ± 0.0050. This is an interactive effect, not evidence that names alone carry the bias. The conclusion requires (i) a random-token masking baseline matched for the number of masked tokens, (ii) ablations that mask stereotypical names while keeping gendered words, and (iii) a significance test across the 15 runs. As reported, the experiments do not isolate names or gendered words as the causal carrier.","section":"Section 4.4.1, Table 4"},{"comment":"The SEAT analysis is difficult to verify and the conclusion that SEAT is not an effective measure does not follow from the displayed numbers. The column header 'BERT Debiased-bert ours SENT-DEBIAS BERT' is malformed; rows such as C6b and C7b appear to show p-values below 0.05 for the debiased model (0.03 and 0.04), yet the text says that 4/6 sets are not statistically significant. Because the paper uses this conclusion to justify abandoning SEAT in favor of the proposed metrics, the table must be unambiguous about which model each effect size and p-value belong to, and the counting must be consistent with the table.","section":"Section 3.4.3, Table 3"},{"comment":"The continued pre-training experiments are not reproducible. There is no report of the BERT variant, number of training steps or epochs, batch size, learning rate, mask probability, the exact attribute-word list, or the weight of the regularizer in Eq. (1). Section 3.3 names BookCorpus but gives no training schedule. Without these details, the reader cannot assess whether the regularizer dominates the masked-LM loss or how much the debiasing costs in model quality. This is especially important because the paper's contextual method is the primary contribution.","section":"Section 3.2.1 and Section 3.3"},{"comment":"The bias-score reductions on CrowS-pairs-style evaluation are reported without any uncertainty or significance testing; for example, Type1 test values 53.54 vs 51.52 and Type2 test values 60.61 vs 54.04 are presented as improvements with no error bars, number of runs, or paired test. Given the known variance of template-based stereotype scores, these differences may not be reliable. The caption and text should also state consistently that lower values are better, since Table 1 already says that higher numbers indicate higher model bias.","section":"Section 3.4.2, Table 1"}],"minor_comments":[{"comment":"There are unresolved cross-references: Section A.1.1 refers to 'sec. ??' and Appendix A.1.2 refers to 'Figure ??(a)' and 'Figure ??(b)', which should be resolved before submission.","section":"Appendix A.1.1 and A.1.2"},{"comment":"There are several typos: 'intuion' in Section 4.4.2, 'debasing' in Section 1, 'smenatics' in Appendix A.3.3, and 'explicity' in the Table 7 caption.","section":"Section 4.4.2, Section 1, Appendix A.3.3, Table 7"},{"comment":"The text mentions 'shown in Assignment 3,' which appears to refer to an external course assignment; this should be removed or replaced with a citation or an appendix section.","section":"Section 3.4.3"},{"comment":"The criterion that target occupations appear 'in the top 30000 predicted words' needs a motivation and a statement of the BERT vocabulary size; otherwise the effective test set is unclear.","section":"Section 3.4.1"},{"comment":"The downstream task results are reported without standard deviations across fine-tuning runs; the claimed 0.8% gain on SST-2 (93.4 vs 92.6) is within typical run-to-run variability for GLUE fine-tuning.","section":"Table 2"}],"recommendation":"reject","confidential_remarks":"The paper appears to be a workshop-style report with unresolved references and an unclear evaluation table, and the central contextual result may largely reflect an evaluation that mirrors the training objective. The static clustering experiment is interesting, but the causal claim about names is not supported by the reported ablations. If the authors add a recovery probe, a random-mask baseline, and full training details, the work could become publishable; as presented, I do not recommend it for this journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper has one genuinely interesting empirical result—NER-masking plus explicit gender encoding drops profession clustering accuracy to 0.5157 in static embeddings—and a new-looking regularized MLM objective for BERT. But the central claim that the objective debiases contextual embeddings is not backed by the evidence as written. The intrinsic evaluation largely measures what the regularizer optimizes, and no probe tests whether gender is still recoverable from unmasked context at inference. I'd send it to review only with the expectation of major revision or rejection; the static result is worth chasing but the contextual half needs real work.\n\nWhat's new: the attention-masked MLM with a logit-difference regularizer (Eq. 1) is not in the cited prior work. The combination of NER-masking and explicit gender encoding for static embeddings is also new, and the near-chance clustering is a clean demonstration that names can carry more bias signal than gendered words. The paper also does the right thing by checking downstream tasks (SST-2, CoLA, QNLI) and by engaging directly with the Gonen–Goldberg critique.\n\nSoft spots, in order of severity. First, the contextual evaluation is partly circular: the main metric—probability of he/she given an occupation, and vice versa—is precisely what the regularizer drives toward equality for gender-neutral words. The paper acknowledges this alignment but doesn't treat it as a limitation. Second, the training-time attention mask doesn't guarantee inference-time debiasing; gender information can route through other tokens. The paper invokes Gonen–Goldberg for static embeddings but runs no probing or recovery experiment for BERT. Third, the BERT continued pre-training is underspecified: no base model, steps, learning rate, loss weight, or compute. Fourth, reproducibility suffers from missing code/data and unfinished placeholders ('Figure ??', 'sec. ??', 'Assignment 3'). Fifth, the 'empirical proof' that names are the main source of static bias lacks a control experiment for general token masking—masking any frequent entity type might break spurious correlations.\n\nThe static experiment is the strongest part, and the NER-M+EGE result deserves follow-up. But as a standalone paper, the contextual claims overreach. I'd recommend that a serious editor send it to referees—there's enough substance to warrant detailed comments—but the outcome should be reject-and-resubmit or major revision, not acceptance.","headline":"The static NER-M+EGE result is the real story here; the BERT debiasing claim is not established by the current evaluation.","tokens_in":15937,"tokens_out":3946,"would_cite":false,"duration_ms":37176,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A training-time objective that blocks gendered words when predicting neutral nouns largely removes gender bias from BERT while preserving downstream performance, and the paper argues that stereotypical names, not gendered words, are the…","keywords":["gender bias","contextual embeddings","masked language modeling","debiasing","BERT","static word embeddings","named entity masking","stereotypical names"],"falsifier":"Train a linear or shallow classifier on the debiased BERT's representations of neutral nouns (for example, 'nurse') to predict the gender of an attribute word or named entity that appears elsewhere in the sentence; if the classifier achieves accuracy well above chance, gender information has survived the attention masking and the debiasing is superficial. For the static-embedding claim, run the Gonen-Goldberg clustering experiment on a corpus with names masked versus a corpus with gendered words removed but names intact; if clustering stays high when names remain, names are indeed the carrier.","tokens_in":14859,"feed_emoji":"⚖️","tokens_out":3792,"duration_ms":37555,"temperature":0.7,"pith_summary":"The paper sets out to show that gender bias in contextual embeddings can be reduced at training time by changing the masked-language-model objective, rather than by post-processing or heavy data augmentation. It proposes an objective in which the model predicts masked nouns without attending to gendered attribute words, and predicts masked attribute words without attending to nouns, plus a regularizer that pushes prediction scores for paired male and female tokens toward equality. The paper also investigates static embeddings and argues that the persistent clustering of stereotyped professions is driven mainly by co-occurring stereotypical names, not by gendered words themselves, and that masking those names while explicitly encoding gender categories brings profession clustering close to chance.","feed_headline":"Attention masks strip gender bias from BERT without hurting accuracy","feed_subtitle":"A training-time objective blocks gendered words when predicting nouns, and names turn out to drive static-embedding bias.","key_machinery":"The central object is a modified masked-language-model training procedure with two attention-mask strategies: when predicting a masked noun, the model is allowed to attend only to non-attribute tokens, and when predicting a masked attribute word, it may attend only to non-noun tokens. A regularizer adds the summed absolute difference of the pre-softmax prediction scores for paired male and female tokens, pushing the model to treat paired gendered words symmetrically for gender-neutral nouns. For static embeddings, the load-bearing mechanisms are an entity-mask corpus (NER-M) that replaces stereotypical male and female names with a token, and an explicit gender-encoding regularizer (EGE) that trains a CBOW model to classify center words as male, female, or neutral; when applied together, they separate professions from name-mediated gender clusters.","core_discovery":"The central claim is that a regularized MLM objective can mitigate gender bias in BERT's contextual embeddings while keeping downstream performance intact, and that the main carrier of gender bias in static embeddings is stereotypical names rather than gendered words. The paper supports the first claim with pronoun-filling and occupation-filling experiments, lower stereotype scores on WinoBias, and retained accuracy on SST-2, CoLA, and QNLI. For static embeddings, it reports that combining NER-Masking of stereotypical names with an explicit gender-encoding regularizer lowers the Gonen-Goldberg profession-clustering accuracy to 0.5157, near chance, whereas removing gendered words alone leaves clustering at 0.5970.","pith_inferences":["If stereotypical names are the primary bias carrier, then masking or rebalancing named entities in training corpora could become a general preprocessing step for other social biases, not just gender.","The contextual method's reliance on attention masks suggests a natural test of its depth: a probing classifier should not be able to recover the gender of a masked attribute word from the remaining context, a test the paper does not run.","The paper's own layer analysis hints that gender information may still be encoded in some BERT layers after debiasing, which would make the contextual method closer to suppression than elimination.","The proposed augmentation strategy, which swaps gendered labels and masks neutral words, may strip so much gender information that the model becomes unable to predict pronouns at all, a side effect observed in the paper's preliminary analysis."],"forward_implications":["BERT-style models can be debiased during continued pre-training without a large drop in downstream performance.","The proposed evaluation metrics give a direct, normatively motivated way to check whether a model is treating paired gendered words symmetrically in context.","For static embeddings, masking stereotypical names appears more effective than removing gendered words, pointing to named entities as the primary bias carrier.","Combining entity masking with explicit gender encoding reduces profession clustering to near chance, a result other debiasing methods do not reach.","Existing post-hoc debiasing methods fail to reduce profession clustering because they do not address the name-mediated transitive path of bias."],"supporting_citations":[{"why":"Supplies the definitional gender pairs, the gender-subspace concept, and the Hard Debias baseline used in the static-embedding experiments.","marker":"(Bolukbasi et al., 2016)"},{"why":"Supplies the profession-clustering experiment and the 'lipstick on a pig' critique that motivates the search for the true carrier of bias.","marker":"(Gonen and Goldberg, 2019)"},{"why":"Provides BERT, the contextual model whose pretraining objective is modified and evaluated.","marker":"(Devlin et al., 2019)"},{"why":"Provides Winogender Schemas and WinoBias, used for the pronoun-filling and occupation-filling evaluations and for the stereotype scores.","marker":"(Rudinger et al., 2018)"},{"why":"Provides the CrowS-pairs stereotype score used to quantify model bias on WinoBias in Table 1.","marker":"(Nangia et al., 2020)"},{"why":"Provides the counterfactual data augmentation baseline for contextual embeddings that the paper extends and critiques.","marker":"(Zhao et al., 2019)"},{"why":"Provides Sent-Debias, the comparison method for SEAT scores and downstream-task performance.","marker":"(Liang et al., 2020)"},{"why":"Shows that stereotypical name pairs estimate the gender direction as effectively as definitional pairs, motivating the NER-Masking approach.","marker":"(Dev and Phillips 2019)"},{"why":"Provides the CBOW/Word2Vec model and original embeddings used for all static-embedding experiments and baselines.","marker":"(Mikolov et al., 2013)"}],"fun_headline_variants":["New MLM objective erases gender bias from BERT, preserves accuracy","Training objective strips gender bias from BERT, not accuracy","Stereotypical names cause static bias; MLM fixes BERT","Regularized MLM debiases BERT, keeps quality on NLP benchmarks","Gender bias in BERT: training tweak removes it, performance intact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The contextual debiasing works only if blocking attention to gendered words prevents the model from encoding gender in the neutral noun's representation; if the remaining context still lets a probe recover the masked word's gender, the bias is hidden, not removed.","fun_headline_variants_meta":{"raw":{"variants":["New MLM objective erases gender bias from BERT, preserves accuracy","Training objective strips gender bias from BERT, not accuracy","Stereotypical names cause static bias; MLM fixes BERT","Regularized MLM debiases BERT, keeps quality on NLP benchmarks","Gender bias in BERT: training tweak removes it, performance intact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001224,"raw_usage":{"total_tokens":5007,"prompt_tokens":895,"completion_tokens":4112,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":4018}},"tokens_in":511,"tokens_out":4112,"duration_ms":30617,"temperature":1.0,"reasoning_tokens":4018,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:56:51.676013+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a linear or shallow classifier on the debiased BERT's representations of neutral nouns (for example, 'nurse') to predict the gender of an attribute word or named entity that appears elsewhere in the sentence; if the classifier achieves accuracy well above chance, gender information has survived the attention masking and the debiasing is superficial. For the static-embedding claim, run the Gonen-Goldberg clustering experiment on a corpus with names masked versus a corpus with gendered words removed but names intact; if clustering stays high when names remain, names are indeed the carrier.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the definitional gender pairs, the gender-subspace concept, and the Hard Debias baseline used in the static-embedding experiments."}],"review_version":1}