Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling

Flavia Salutari; Jerome Ramos; Hossein A. Rahmani; Leonardo Linguaglossa; Aldo Lipani

doi:10.1007/978-3-031-33374-3_42

Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling

Flavia Salutari
, Jerome Ramos
, Hossein A. Rahmani
, Leonardo Linguaglossa
, Aldo Lipani

University College London
Telecom Paris

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

6 Citations (Scopus)

Abstract

In recent years, groundbreaking transformer-based language models (LMs) have made tremendous advances in natural language processing (NLP) tasks. However, the measurement of their fairness with respect to different social groups still remains unsolved. In this paper, we propose and thoroughly validate an evaluation technique to assess the quality and bias of language model predictions on transcripts of both spoken African American English (AAE) and Spoken American English (SAE). Our analysis reveals the presence of a bias towards SAE encoded by state-of-the-art LMs such as BERT and DistilBERT and a lower bias in distilled LMs. We also observe a bias towards AAE in RoBERTa and BART. Additionally, we show evidence that this disparity is present across all the LMs when we only consider the grammar and the syntax specific to AAE.

Original language	English
Title of host publication	Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings
Editors	Hisashi Kashima, Tsuyoshi Ide, Wen-Chih Peng
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	532-543
Number of pages	12
ISBN (Print)	9783031333736
DOIs	https://doi.org/10.1007/978-3-031-33374-3_42
Publication status	Published - 1 Jan 2023
Event	27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023 - Hybrid, Osaka, Japan Duration: 25 May 2023 → 28 May 2023

Publication series

Name	Lecture Notes in Computer Science
Volume	13935 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023
Country/Territory	Japan
City	Hybrid, Osaka
Period	25/05/23 → 28/05/23

Keywords

Bias and Fairness
Evaluation
Language Model
Transformers

Access to Document

10.1007/978-3-031-33374-3_42

Cite this

Salutari, F., Ramos, J., Rahmani, H. A., Linguaglossa, L., & Lipani, A. (2023). Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling. In H. Kashima, T. Ide, & W.-C. Peng (Eds.), Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings (pp. 532-543). (Lecture Notes in Computer Science; Vol. 13935 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-031-33374-3_42

Salutari, Flavia ; Ramos, Jerome ; Rahmani, Hossein A. et al. / Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling. Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings. editor / Hisashi Kashima ; Tsuyoshi Ide ; Wen-Chih Peng. Springer Science and Business Media Deutschland GmbH, 2023. pp. 532-543 (Lecture Notes in Computer Science).

@inproceedings{d39bfeb980c24b3b82098af4daac3e6c,

title = "Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling",

abstract = "In recent years, groundbreaking transformer-based language models (LMs) have made tremendous advances in natural language processing (NLP) tasks. However, the measurement of their fairness with respect to different social groups still remains unsolved. In this paper, we propose and thoroughly validate an evaluation technique to assess the quality and bias of language model predictions on transcripts of both spoken African American English (AAE) and Spoken American English (SAE). Our analysis reveals the presence of a bias towards SAE encoded by state-of-the-art LMs such as BERT and DistilBERT and a lower bias in distilled LMs. We also observe a bias towards AAE in RoBERTa and BART. Additionally, we show evidence that this disparity is present across all the LMs when we only consider the grammar and the syntax specific to AAE.",

keywords = "Bias and Fairness, Evaluation, Language Model, Transformers",

author = "Flavia Salutari and Jerome Ramos and Rahmani, \{Hossein A.\} and Leonardo Linguaglossa and Aldo Lipani",

note = "Publisher Copyright: {\textcopyright} 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.; 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023 ; Conference date: 25-05-2023 Through 28-05-2023",

year = "2023",

month = jan,

day = "1",

doi = "10.1007/978-3-031-33374-3\_42",

language = "English",

isbn = "9783031333736",

series = "Lecture Notes in Computer Science",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "532--543",

editor = "Hisashi Kashima and Tsuyoshi Ide and Wen-Chih Peng",

booktitle = "Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings",

}

Salutari, F, Ramos, J, Rahmani, HA, Linguaglossa, L & Lipani, A 2023, Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling. in H Kashima, T Ide & W-C Peng (eds), Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings. Lecture Notes in Computer Science, vol. 13935 LNCS, Springer Science and Business Media Deutschland GmbH, pp. 532-543, 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Hybrid, Osaka, Japan, 25/05/23. https://doi.org/10.1007/978-3-031-33374-3_42

Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling. / Salutari, Flavia; Ramos, Jerome; Rahmani, Hossein A. et al.
Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings. ed. / Hisashi Kashima; Tsuyoshi Ide; Wen-Chih Peng. Springer Science and Business Media Deutschland GmbH, 2023. p. 532-543 (Lecture Notes in Computer Science; Vol. 13935 LNCS).

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

TY - GEN

T1 - Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling

AU - Salutari, Flavia

AU - Ramos, Jerome

AU - Rahmani, Hossein A.

AU - Linguaglossa, Leonardo

AU - Lipani, Aldo

PY - 2023/1/1

Y1 - 2023/1/1

N2 - In recent years, groundbreaking transformer-based language models (LMs) have made tremendous advances in natural language processing (NLP) tasks. However, the measurement of their fairness with respect to different social groups still remains unsolved. In this paper, we propose and thoroughly validate an evaluation technique to assess the quality and bias of language model predictions on transcripts of both spoken African American English (AAE) and Spoken American English (SAE). Our analysis reveals the presence of a bias towards SAE encoded by state-of-the-art LMs such as BERT and DistilBERT and a lower bias in distilled LMs. We also observe a bias towards AAE in RoBERTa and BART. Additionally, we show evidence that this disparity is present across all the LMs when we only consider the grammar and the syntax specific to AAE.

AB - In recent years, groundbreaking transformer-based language models (LMs) have made tremendous advances in natural language processing (NLP) tasks. However, the measurement of their fairness with respect to different social groups still remains unsolved. In this paper, we propose and thoroughly validate an evaluation technique to assess the quality and bias of language model predictions on transcripts of both spoken African American English (AAE) and Spoken American English (SAE). Our analysis reveals the presence of a bias towards SAE encoded by state-of-the-art LMs such as BERT and DistilBERT and a lower bias in distilled LMs. We also observe a bias towards AAE in RoBERTa and BART. Additionally, we show evidence that this disparity is present across all the LMs when we only consider the grammar and the syntax specific to AAE.

KW - Bias and Fairness

KW - Evaluation

KW - Language Model

KW - Transformers

U2 - 10.1007/978-3-031-33374-3_42

DO - 10.1007/978-3-031-33374-3_42

M3 - Conference contribution

AN - SCOPUS:85173561010

SN - 9783031333736

T3 - Lecture Notes in Computer Science

SP - 532

EP - 543

BT - Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings

A2 - Kashima, Hisashi

A2 - Ide, Tsuyoshi

A2 - Peng, Wen-Chih

PB - Springer Science and Business Media Deutschland GmbH

T2 - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023

Y2 - 25 May 2023 through 28 May 2023

ER -

Salutari F, Ramos J, Rahmani HA, Linguaglossa L, Lipani A. Quantifying the Bias of Transformer-Based Language Models for African American English in Masked Language Modeling. In Kashima H, Ide T, Peng WC, editors, Advances in Knowledge Discovery and Data Mining - 27th Pacific-Asia Conference on Knowledge Discovery and Data Mining, PAKDD 2023, Proceedings. Springer Science and Business Media Deutschland GmbH. 2023. p. 532-543. (Lecture Notes in Computer Science). doi: 10.1007/978-3-031-33374-3_42