A robust bias mitigation procedure based on the Stereotype Content model

Eddie Ungless; Amy Rafferty; Hrichika Nag; Björn Ross

A robust bias mitigation procedure based on the Stereotype Content model

Eddie Ungless, Amy Rafferty, Hrichika Nag, Björn Ross

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

Abstract

The Stereotype Content model (SCM) states that we tend to perceive minority groups as cold, incompetent or both. In this paper we adapt existing work to demonstrate that the Stereotype Content model holds for contextualised word embeddings, then use these results to evaluate a fine-tuning process designed to drive a language model away from stereotyped portrayals of minority groups. We find the SCM terms are better able to capture bias than demographic agnostic terms related to pleasantness. Further, we were able to reduce the presence of stereotypes in the model through a simple fine-tuning procedure that required minimal human and computer resources, without harming downstream performance. We present this work as a prototype of a debiasing procedure that aims to remove the need for a priori knowledge of the specifics of bias in the model.

Original language	English
Title of host publication	Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science
Editors	David Bamman, Dirk Hovy, David Jurgens, Katherine Keith, Brendan O'Connor, Svitlana Volkova
Publisher	Association for Computational Linguistics (ACL)
Pages	207-217
Number of pages	11
ISBN (Print)	978-1-959429-20-3
Publication status	Published - 2 Feb 2023
Event	The 5th Workshop on Natural Language Processing and Computational Social Science - Abu Dhabi, United Arab Emirates Duration: 7 Dec 2022 → 7 Dec 2022 Conference number: 5 https://sites.google.com/site/nlpandcss/home/nlp-css-at-emnlp-2022

Workshop

Workshop	The 5th Workshop on Natural Language Processing and Computational Social Science
Abbreviated title	NLP+CSS 2022
Country/Territory	United Arab Emirates
City	Abu Dhabi
Period	7/12/22 → 7/12/22
Internet address	https://sites.google.com/site/nlpandcss/home/nlp-css-at-emnlp-2022

Access to Document

A Robust Bias_UNGLESS_DOA15102022_VOR_CC-BYFinal published version, 263 KBLicence: Creative Commons: Attribution (CC-BY)

Cite this

Ungless, E., Rafferty, A., Nag, H., & Ross, B. (2023). A robust bias mitigation procedure based on the Stereotype Content model. In D. Bamman, D. Hovy, D. Jurgens, K. Keith, B. O'Connor, & S. Volkova (Eds.), Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science (pp. 207-217). Association for Computational Linguistics (ACL). https://arxiv.org/abs/2210.14552v1

Ungless, Eddie ; Rafferty, Amy ; Nag, Hrichika et al. / A robust bias mitigation procedure based on the Stereotype Content model. Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science. editor / David Bamman ; Dirk Hovy ; David Jurgens ; Katherine Keith ; Brendan O'Connor ; Svitlana Volkova. Association for Computational Linguistics (ACL), 2023. pp. 207-217

@inproceedings{4972924ac6b14566a3500c66f69d6940,

title = "A robust bias mitigation procedure based on the Stereotype Content model",

abstract = "The Stereotype Content model (SCM) states that we tend to perceive minority groups as cold, incompetent or both. In this paper we adapt existing work to demonstrate that the Stereotype Content model holds for contextualised word embeddings, then use these results to evaluate a fine-tuning process designed to drive a language model away from stereotyped portrayals of minority groups. We find the SCM terms are better able to capture bias than demographic agnostic terms related to pleasantness. Further, we were able to reduce the presence of stereotypes in the model through a simple fine-tuning procedure that required minimal human and computer resources, without harming downstream performance. We present this work as a prototype of a debiasing procedure that aims to remove the need for a priori knowledge of the specifics of bias in the model.",

author = "Eddie Ungless and Amy Rafferty and Hrichika Nag and Bj{\"o}rn Ross",

note = "We would like to thank our anonymous reviewers for their feedback. This work is in part supported by the UKRI Centre for Doctoral Training in Natural Language Processing, funded by the UKRI (grant EP/S022481/1) and the University of Edinburgh, School of Informatics; The 5th Workshop on Natural Language Processing and Computational Social Science, NLP+CSS 2022 ; Conference date: 07-12-2022 Through 07-12-2022",

year = "2023",

month = feb,

day = "2",

language = "English",

isbn = "978-1-959429-20-3",

pages = "207--217",

editor = "David Bamman and Dirk Hovy and David Jurgens and Katherine Keith and Brendan O'Connor and Svitlana Volkova",

booktitle = "Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science",

publisher = "Association for Computational Linguistics (ACL)",

url = "https://sites.google.com/site/nlpandcss/home/nlp-css-at-emnlp-2022",

}

Ungless, E, Rafferty, A, Nag, H & Ross, B 2023, A robust bias mitigation procedure based on the Stereotype Content model. in D Bamman, D Hovy, D Jurgens, K Keith, B O'Connor & S Volkova (eds), Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science. Association for Computational Linguistics (ACL), pp. 207-217, The 5th Workshop on Natural Language Processing and Computational Social Science, Abu Dhabi, United Arab Emirates, 7/12/22. <https://arxiv.org/abs/2210.14552v1>

A robust bias mitigation procedure based on the Stereotype Content model. / Ungless, Eddie; Rafferty, Amy; Nag, Hrichika et al.

Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science. ed. / David Bamman; Dirk Hovy; David Jurgens; Katherine Keith; Brendan O'Connor; Svitlana Volkova. Association for Computational Linguistics (ACL), 2023. p. 207-217.

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - A robust bias mitigation procedure based on the Stereotype Content model

AU - Ungless, Eddie

AU - Rafferty, Amy

AU - Nag, Hrichika

AU - Ross, Björn

N1 - Conference code: 5

PY - 2023/2/2

Y1 - 2023/2/2

N2 - The Stereotype Content model (SCM) states that we tend to perceive minority groups as cold, incompetent or both. In this paper we adapt existing work to demonstrate that the Stereotype Content model holds for contextualised word embeddings, then use these results to evaluate a fine-tuning process designed to drive a language model away from stereotyped portrayals of minority groups. We find the SCM terms are better able to capture bias than demographic agnostic terms related to pleasantness. Further, we were able to reduce the presence of stereotypes in the model through a simple fine-tuning procedure that required minimal human and computer resources, without harming downstream performance. We present this work as a prototype of a debiasing procedure that aims to remove the need for a priori knowledge of the specifics of bias in the model.

AB - The Stereotype Content model (SCM) states that we tend to perceive minority groups as cold, incompetent or both. In this paper we adapt existing work to demonstrate that the Stereotype Content model holds for contextualised word embeddings, then use these results to evaluate a fine-tuning process designed to drive a language model away from stereotyped portrayals of minority groups. We find the SCM terms are better able to capture bias than demographic agnostic terms related to pleasantness. Further, we were able to reduce the presence of stereotypes in the model through a simple fine-tuning procedure that required minimal human and computer resources, without harming downstream performance. We present this work as a prototype of a debiasing procedure that aims to remove the need for a priori knowledge of the specifics of bias in the model.

M3 - Conference contribution

SN - 978-1-959429-20-3

SP - 207

EP - 217

BT - Proceedings of the Fifth Workshop on Natural Language Processing and Computational Social Science

A2 - Bamman, David

A2 - Hovy, Dirk

A2 - Jurgens, David

A2 - Keith, Katherine

A2 - O'Connor, Brendan

A2 - Volkova, Svitlana

PB - Association for Computational Linguistics (ACL)

T2 - The 5th Workshop on Natural Language Processing and Computational Social Science

Y2 - 7 December 2022 through 7 December 2022

ER -

A robust bias mitigation procedure based on the Stereotype Content model

Abstract

Workshop

Access to Document

Fingerprint

Cite this