Towards Debiasing Sentence Representations

Liang, Paul Pu; Li, Irene Mengze; Zheng, Emily; Lim, Yao Chong; Salakhutdinov, Ruslan; Morency, Louis–Philippe

doi:10.18653/v1/2020.acl-main.488

Cited by 85 publications

(89 citation statements)

References 46 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Despite the relatively ample literature on gender debiasing for word-level representations, very little work has focused on sentence representations (Liang et al, 2020;Liu et al, 2019;Lee et al, 2019b). Until this point, most debiasing work on sentences mainly focus on measuring bias (Lee et al, 2019b;.…”

Section: Related Workmentioning

confidence: 99%

Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation

Dinan

Fan

Williams

et al. 2020

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

105

View full text Add to dashboard Cite

Social biases present in data are often directly reflected in the predictions of models trained on that data. We analyze gender bias in dialogue data, and examine how this bias is not only replicated, but is also amplified in subsequent generative chit-chat dialogue models. We measure gender bias in six existing dialogue datasets before selecting the most biased one, the multi-player textbased fantasy adventure dataset LIGHT (Urbanek et al., 2019), as a testbed for bias mitigation techniques. We consider three techniques to mitigate gender bias: counterfactual data augmentation, targeted data collection, and bias controlled training. We show that our proposed techniques mitigate gender bias by balancing the genderedness of generated dialogue utterances, and find that they are particularly effective in combination. We evaluate model performance with a variety of quantitative methods-including the quantity of gendered words, a dialogue safety classifier, and human assessments-all of which show that our models generate less gendered, but equally engaging chit-chat responses.

show abstract

Section: Related Workmentioning

confidence: 99%

Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation

Dinan

Fan

Williams

et al. 2020

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

105

View full text Add to dashboard Cite

show abstract

“…To sum up, people reveal their stereotype expectancies in many subtle ways in the words they use. This fact can explain the effectiveness of several computational works at measuring social bias (e.g., gender, racial, religion, and ethnic stereotypes among others) by using word representations [28,41,42]. In [41], social bias are quantified by using embeddings of representative words such as women, men, Asians living in the United States, and white people (i.e., non-Hispanic subpopulation from the United States).…”

Section: Related Workmentioning

confidence: 99%

“…In [28,42], word embeddings are used to measure (and reduce) the bias. In [28], a methodology based directly on word embeddings is proposed to differentiate gender bias associations (e.g., biased association between receptionist and female) from associations of related concepts (e.g., between queen and female); a neutralisation and equalisation debias process.…”

Section: Related Workmentioning

confidence: 99%

“…In [28], a methodology based directly on word embeddings is proposed to differentiate gender bias associations (e.g., biased association between receptionist and female) from associations of related concepts (e.g., between queen and female); a neutralisation and equalisation debias process. The same debias process is used in [42], but in this work a new representation to detect the bias is proposed: the authors include a contextualisation step and create two subspaces of the attribute (e.g., gender), one for male and other for female. The contextualisation relies on a large and diverse set of sentences in which the bias attribute words (e.g., he/she, man/woman) appear; for example, if in one subspace it is included a sentence containing he, the same sentence is included in the subspace for female but replacing he by she.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

How Do You Speak about Immigrants? Taxonomy and StereoImmigrants Dataset for Identifying Stereotypes about Immigrants

et al. 2021

View full text Add to dashboard Cite

Stereotype is a type of social bias massively present in texts that computational models use. There are stereotypes that present special difficulties because they do not rely on personal attributes. This is the case of stereotypes about immigrants, a social category that is a preferred target of hate speech and discrimination. We propose a new approach to detect stereotypes about immigrants in texts focusing not on the personal attributes assigned to the minority but in the frames, that is, the narrative scenarios, in which the group is placed in public speeches. We have proposed a fine-grained social psychology grounded taxonomy with six categories to capture the different dimensions of the stereotype (positive vs. negative) and annotated a novel StereoImmigrants dataset with sentences that Spanish politicians have stated in the Congress of Deputies. We aggregate these categories in two supracategories: one is Victims that expresses the positive stereotypes about immigrants and the other is Threat that expresses the negative stereotype. We carried out two preliminary experiments: first, to evaluate the automatic detection of stereotypes; and second, to distinguish between the two supracategories of immigrants’ stereotypes. In these experiments, we employed state-of-the-art transformer models (monolingual and multilingual) and four classical machine learning classifiers. We achieve above 0.83 of accuracy with the BETO model in both experiments, showing that transformers can capture stereotypes about immigrants with a high level of accuracy.

show abstract

“…In four tasks with known bias factors -hate speech detection, toxicity detection, occupation prediction from short bios, and coreference resolution -we explore whether upstream bias mitigation of a LM followed by downstream fine-tuning reduces bias for the downstream model. Though previous work has addressed biases in frozen PTLM or word embeddings (Bolukbasi et al, 2016;Zhou et al, 2019;Bhardwaj et al, 2020;Liang et al, 2020;Ravfogel et al, 2020), for example by measuring associations between gender and occupations in an embedding space, they do not study their effect on downstream classifiers (Fig. 1 (b)), while some of them study the effects while keeping the embeddings frozen Kurita et al, 2019;Prost et al, 2019).…”

Section: Introductionmentioning

confidence: 99%

On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning

Jin¹,

Barbieri²,

Kennedy³

et al. 2021

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langua

View full text Add to dashboard Cite

Fine-tuned language models have been shown to exhibit biases against protected groups in a host of modeling tasks such as text classification and coreference resolution. Previous works focus on detecting these biases, reducing bias in data representations, and using auxiliary training objectives to mitigate bias during fine-tuning. Although these techniques achieve bias reduction for the task and domain at hand, the effects of bias mitigation may not directly transfer to new tasks, requiring additional data collection and customized annotation of sensitive attributes, and re-evaluation of appropriate fairness metrics. We explore the feasibility and benefits of upstream bias mitigation (UBM) for reducing bias on downstream tasks, by first applying bias mitigation to an upstream model through fine-tuning and subsequently using it for downstream finetuning. We find, in extensive experiments across hate speech detection, toxicity detection, occupation prediction, and coreference resolution tasks over various bias factors, that the effects of UBM are indeed transferable to new downstream tasks or domains via finetuning, creating less biased downstream models than directly fine-tuning on the downstream task or transferring from a vanilla upstream model. Though challenges remain, we show that UBM promises more efficient and accessible bias mitigation in LM fine-tuning. 12

show abstract

Towards Debiasing Sentence Representations

Cited by 85 publications

References 46 publications

Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation

Queens are Powerful too: Mitigating Gender Bias in Dialogue Generation

How Do You Speak about Immigrants? Taxonomy and StereoImmigrants Dataset for Identifying Stereotypes about Immigrants

On Transferability of Bias Mitigation Effects in Language Model Fine-Tuning

Contact Info

Product

Resources

About