A Dutch BERT model encodes gender linearly by epoch 20 but does not dynamically update its representations when explicit female cues contradict learned stereotypical associations in short sentence templates.
Science , volume=
3 Pith papers cite this work. Polarity classification is still indexing.
3
Pith papers citing it
fields
cs.CL 3years
2026 3verdicts
CONDITIONAL 3representative citing papers
Debiasing language-model training data for a target group frequently increases stereotyping or counter-stereotyping for non-target groups across categories, models, and scales.
Unsupervised embedding proxies for social constructs are non-identified mixtures of target and confounders, so the Construct Validity Protocol plus Counterfactual Neutralization are required to turn them into defensible measures.
citing papers explorer
-
Is She Even Relevant? When BERT Ignores Explicit Gender Cues
A Dutch BERT model encodes gender linearly by epoch 20 but does not dynamically update its representations when explicit female cues contradict learned stereotypical associations in short sentence templates.
-
When Debiasing Backfires: Counterintuitive Side Effects of Preprocessing-Based Stereotype Mitigation
Debiasing language-model training data for a target group frequently increases stereotyping or counter-stereotyping for non-target groups across categories, models, and scales.
-
The Proxy Presumption: From Semantic Embeddings to Valid Social Measures
Unsupervised embedding proxies for social constructs are non-identified mixtures of target and confounders, so the Construct Validity Protocol plus Counterfactual Neutralization are required to turn them into defensible measures.