Lipstick on a Pig: Debiasing Methods Cover up Systematic Gender Biases in Word Embeddings But do not Remove Them

By Hila Gonen, Yoav Goldberg. 2019

47 citations recorded by the source corpus.

Subfields: Natural Language Processing, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Natural Language Processing

In Safety, Robustness & Interpretability

Citations in the corpus

Reducing Sentiment Bias in Language Models via Counterfactual Evaluation (2019, background): “Gonen & Goldberg (2019) however pointed out limitations to this approach: bias in word embeddings appear indirectly in other ways, even after minimizing linear projections onto gender-related subspaces.”
Fair Is Better than Sensational: Man Is to Doctor as Woman Is to Doctor (2019, background): “If analogies might not be the most appropriate tool to capture certain relations, surely matters have been made worse by the way that consciously or not they have been used (Gonen and Goldberg [2019] have rightly dubbed them sensational “party tricks”).”
SuperGLUE: A Stickier Benchmark for General-Purpose Language Understanding Systems (2019, neutral): “(Gonen and Goldberg, 2019).”
Toward Gender-Inclusive Coreference Resolution (2019, background): “…in coreference resolution, including, for instance, natural language inference [Rudinger et al. 2017], word embeddings [e.g., Bolukbasi et al. 2016; Gonen and Goldberg 2019; Romanov et al. 2019], sentiment analysis [Kir-itchenko and Mohammad 2018], machine translation [Font and Costa-jussà 2019;…”
On Measuring and Mitigating Biased Inferences of Word Embeddings (2019, background): “This simple approach thus, avoids the trap of residual information (Gonen and Goldberg 2019) seen in hard debiasing approach of (Bolukbasi et al. 2016), which categorizes words and treats each category differently.”
Queens Are Powerful Too: Mitigating Gender Bias in Dialogue Generation (2019, methodology): “they have only succeeded in hiding word embedding biases as opposed to removing them (Gonen and Goldberg, 2019)—making gender debiasing still an open area of research.”
Identifying and Reducing Gender Bias in Word-Level Language Models (2019, methodology): “It can be concluded different bias metrics show different kinds of bias (Gonen and Goldberg, 2019).”
Large image datasets: A pyrrhic win for computer vision? (2020, background): “The idea of a complete removal of biases, for example, might in reality be simply hiding them out of sight [36].”
Hurtful words: quantifying biases in clinical contextual word embeddings (2020, background): “As shown here and in other work [19, 24], debiasing methods might not always be effective and might simply hide the bias instead of removing it [24].”
Reducing sentiment polarity for demographic attributes in word embeddings using adversarial learning (2020, background): “However, identifying and mitigating unintended bias is still a vast and complicated problem as [11] argues.”
RobBERT: a Dutch RoBERTa-based Language Model (2020, unknown): “, 2017) or embedding analysis (Gonen and Goldberg, 2019).”
The Importance of Modeling Social Factors of Language: Theory and Practice (2021, methodology): “For that, we bor-row from Grice (1975), who laid out four maxims that govern effective communication in social situations.”
The practical ethics of bias reduction in machine translation: why domain adaptation is better than data debiasing (2021, background): “However, a subsequent analysis by Gonen and Goldberg (2019) has argued that such methods are merely ‘lipstick on a pig’ – that is, while the gender bias ‘direction’ of the word embeddings is superficially reduced, relationships between gendered terms learned from the data are still detec – and thus…”
Debiasing Word Embeddings from Sentiment Associations in Names (2020, background): “However, a recent study by Gonen and Goldberg [9] shows that both, the post-processing and the isolation approach, remove bias only at a superﬁcial level.”
Bias in machine learning - what is it good for? (2020, methodology): “However, bias may still remain [18] after applying these methods, and may propagate to models generated by other machine learning algorithms that rely on word embeddings as input.”
A Review on Fairness in Machine Learning (2022, background): “Curiously, a recent study has argued that a major concern is that some of the proposed methods for removing biases in word embeddings are actually not able to remove biases but rather just “hide” them [85].”
Investigating Gender Bias in BERT (2020, unknown): “n of possible gender subspace in BERT from (Bolukbasi et al. 2016), which studied the existence of gender directions in static word embeddings. (Zhao et al. 2019; Basta, Costa-juss`a, and Casas 2019; Gonen and Goldberg 2019) focused their study on ELMo. (Kurita et al. 2019) provided a template-based”