Adversarial Removal of Demographic Attributes from Text Data.

The implication is that decisions of classifiers trained on textual data are not agnostic to -- and likely condition on -- demographic attributes. ... We show that demographic information of authors is encoded in -- and can be recovered from -- the intermediate representations learned by text-based neural classifiers. ... Recent work on creating private representation in the text domain (Li et al., 2018) share our motivation of removing unintended demographic attributes from the learned representation using adversarial ...

arXiv:1808.06640v2 fatcat:j6vrcvpvajhdxpzzgwjlaal75q

Multiple Versions

The implication is that decisions of classifiers trained on textual data are not agnostic to-and likely condition on-demographic attributes. ... We show that demographic information of authors is encoded in-and can be recovered from-the intermediate representations learned by text-based neural classifiers. ... Recent work on creating private representation in the text domain (Li et al., 2018) share our motivation of removing unintended demographic attributes from the learned representation using adversarial ...

doi:10.18653/v1/d18-1002 dblp:conf/emnlp/ElazarG18 fatcat:6j7hkybrq5aftcicdcwe7ivovm

Contextual representations learned by language models can often encode undesirable attributes, like demographic associations of the users, while being trained for an unrelated target task. ... In this paper, we present an adversarial learning framework "Adversarial Scrubber" (ADS), to debias contextual representations. ... In this work, we focus on removing demographic attributes encoded in data representations during training text classification systems. ...

arXiv:2109.08613v1 fatcat:evjk4v3ubzaljnl5p65ypcjjda

This bias in annotated training data and the tendency of machine learning models to amplify it cause AAE text to often be mislabeled as abusive/offensive/hate speech with a high false positive rate by ... In this paper, we use adversarial training to mitigate this bias, introducing a hate speech classifier that learns to detect toxic sentences while demoting confounds corresponding to AAE texts. ... Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF. ...

arXiv:2005.12246v1 fatcat:klt4rbqn3fbtrnlpyjkzuw5nyu

In this paper, we experiment with text classification of the correlated attributes of document topic and author gender, using a novel multilingual parallel corpus of TED talk transcripts. ... In social media analysis, this problem surfaces for demographic user classes such as language, topic, or gender, which influence how an author writes a text to a substantial extent. ... Elazar and Goldberg (2018) apply the idea to the removal of demographic bias; McHardy et al. (2019) remove publication source as a bias variable from a satire detection model. reported that adversarial ...

dblp:conf/wassa/DayanikP21 fatcat:2kjbzfu3ejes3ghdxplrdmf2n4

In this study, we examine empirically how text data can vary across four demographic factors: gender, age, country, and region. ... We propose a multitask neural model to account for demographic variations via adversarial training. ... The authors thank Zijiao Yang for helping evaluate inference accuracy of the Microsoft Face API. This work was supported in part by the National Science Foundation under award number IIS-1657338. ...

doi:10.18653/v1/s19-1015 dblp:conf/starsem/HuangP19 fatcat:pl2vtoywsnbzzgyrrmsibcxrfq

Instead of relying on unknown decision systems or human decisionmakers, we argue that a better way to protect data providers is to remove the trails of sensitive information before publishing the data. ... Recent work from the NLP community focuses on building systems that make fair decisions based on text. ... Related Work Achieving fairness or preserving privacy through removing sensitive information from text has been explored by adversarial training (Li et al., 2018; Figure 2: Sample of original text, ...

doi:10.18653/v1/w19-8633 dblp:conf/inlg/XuQXC19 fatcat:oxjgrkqctza7xdivajsx2f4iuy

Written text often provides sufficient clues to identify the author, their gender, age, and other important attributes. ... In this paper, we propose an approach to explicitly obscure important author characteristics at training time, such that representations learned are invariant to these attributes. ... expect a larger impact from bias removal using ADV training. ...

arXiv:1805.06093v1 fatcat:pyc3hmzl7vaxlcpwfgpwjkykay

Written text often provides sufficient clues to identify the author, their gender, age, and other important attributes. ... In this paper, we propose an approach to explicitly obscure important author characteristics at training time, such that representations learned are invariant to these attributes. ... expect a larger impact from bias removal using ADV training. ...

doi:10.18653/v1/p18-2005 dblp:conf/acl/LiBC18 fatcat:3hr32nulofcb5hrxm4pqd2mlwi

We extend the method in the form of a gated model, which incorporates protected attributes as input, and show that it is effective at reducing bias in predictions through demographic input perturbation ... Group bias in natural language processing tasks manifests as disparities in system error rates across texts authorized by different demographic groups, typically disadvantaging minority groups. ... The baseline models are: STANDARD, which is a naively-trained MLP classifier; INLP (Ravfogel et al., 2020) , which removes demographic information from text representations through iterative nullspace ...

arXiv:2109.08253v2 fatcat:vdtzadpf3jhifmeljk62xtkpyi

Multiple Versions

We provide a review covering recent progresses to tackle algorithmic fairness problems of deep learning from the computational perspective. ... We also discuss fairness mitigation approaches categorized according to three stages of deep learning life-cycle, aiming to push forward the area of fairness in deep learning and build genuinely fair and ... In-processing Adversarial Learning: From model training perspective, adversarial training [45] is a representative solution to remove information about sensitive attributes from intermediate representation ...

arXiv:1908.08843v2 fatcat:kaaevm64fbctpjfdycv5uz3dhi

Multiple Versions

We address both of these challenges in this paper: First, we evaluate different bias measures and propose the use of retrieval metrics to image-text representations via a bias measuring framework. ... Prior proposed bias measurements lack robustness and feature degradation occurs when mitigating bias without access to pretraining data. ... unbiased, defined as outputting similar distributions of scores across attributes for a given text query which should be unrelated to demographic affiliation. ...

arXiv:2203.11933v2 fatcat:d5ahgqbrmbhlndyr6ln5tjvbwa

Open Access Multiple Versions

Further, we discuss open challenges in privacy-preserving NLP regarding data traceability, computation overhead, dataset size, the prevalence of human biases in embeddings, and the privacy-utility tradeoff ... Data protection laws, such as the European Union's General Data Protection Regulation (GDPR), thereby enforce the need for privacy. ... Reducing or removing human biases from embedding models mainly relies on adversarial training. ...

arXiv:2205.10095v1 fatcat:rksy7oxxlbde5bol3ay44yycru

Open Access

The vast availability of textual data on social media has led to an interest in algorithms to predict user attributes such as gender based on the user's writing. ... away their demographic identities. ... Paraphrasing algorithms benefit from parallel data: texts expressing the same message written by users from different demographic groups. ...

doi:10.18653/v1/w16-5603 dblp:conf/acl-nlpcss/ReddyK16 fatcat:2fmco776fvacnemldd77vxmkle

We apply adversarial learning to both of them to reduce the biases brought by sensitive user attributes. ... However, these models can easily inherit the biases related to users' sensitive attributes (e.g., demographics) encoded in training click data, and may generate recommendation results that are unfair to ... In fact, click behavior data inherently encodes biases related to users' sensitive attributes such as demographics [1, 5] . ...

arXiv:2204.00541v1 fatcat:ludwigtdeffdbfjzxfjl55rbwa

Open Access

Adversarial Removal of Demographic Attributes from Text Data [article]

Preserved Fulltext

Other Versions

Adversarial Removal of Demographic Attributes from Text Data

Preserved Fulltext

Adversarial Scrubbing of Demographic Information for Text Classification [article]

Preserved Fulltext

Demoting Racial Bias in Hate Speech Detection [article]

Preserved Fulltext

Disentangling Document Topic and Author Gender in Multiple Languages: Lessons for Adversarial Debiasing

Preserved Fulltext

Neural User Factor Adaptation for Text Classification: Learning to Generalize Across Author Demographics

Preserved Fulltext

Privacy-Aware Text Rewriting

Preserved Fulltext

Towards Robust and Privacy-preserving Text Representations [article]

Preserved Fulltext

Towards Robust and Privacy-preserving Text Representations

Preserved Fulltext

Balancing out Bias: Achieving Fairness Through Balanced Training [article]

Preserved Fulltext

Other Versions

Fairness in Deep Learning: A Computational Perspective [article]

Preserved Fulltext

Other Versions

A Prompt Array Keeps the Bias Away: Debiasing Vision-Language Models with Adversarial Learning [article]

Preserved Fulltext

Other Versions

How to keep text private? A systematic review of deep learning methods for privacy-preserving natural language processing [article]

Preserved Fulltext

Obfuscating Gender in Social Media Writing

Preserved Fulltext

FairRank: Fairness-aware Single-tower Ranking Framework for News Recommendation [article]

Preserved Fulltext