Gendered Mental Health Stigma in Masked Language Models Inna Wanyin Lin1Lucille Njoo1Anjalie Field2Ashish Sharma1 Katharina Reinecke1Tim Althoff1Yulia Tsvetkov1
2025-05-06
0
0
842.65KB
20 页
10玖币
侵权投诉
Gendered Mental Health Stigma in Masked Language Models
Inna Wanyin Lin1∗Lucille Njoo1∗Anjalie Field2Ashish Sharma1
Katharina Reinecke1Tim Althoff1Yulia Tsvetkov1
1Paul G. Allen School of Computer Science & Engineering, University of Washington
2Stanford University
{ilin, lnjoo}@cs.washington.edu
Abstract
Mental health stigma prevents many individ-
uals from receiving the appropriate care, and
social psychology studies have shown that
mental health tends to be overlooked in men.
In this work, we investigate gendered men-
tal health stigma in masked language mod-
els. In doing so, we operationalize men-
tal health stigma by developing a framework
grounded in psychology research: we use clin-
ical psychology literature to curate prompts,
then evaluate the models’ propensity to gen-
erate gendered words. We find that masked
language models capture societal stigma about
gender in mental health: models are consis-
tently more likely to predict female subjects
than male in sentences about having a men-
tal health condition (32% vs. 19%), and this
disparity is exacerbated for sentences that indi-
cate treatment-seeking behavior. Furthermore,
we find that different models capture dimen-
sions of stigma differently for men and women,
associating stereotypes like anger, blame, and
pity more with women with mental health con-
ditions than with men. In showing the complex
nuances of models’ gendered mental health
stigma, we demonstrate that context and over-
lapping dimensions of identity are important
considerations when assessing computational
models’ social biases.
1 Introduction
Mental health issues are heavily stigmatized, pre-
venting many individuals from seeking appropri-
ate care (Sickel et al.,2014). In addition, social
psychology studies have shown that this stigma
manifests differently for different genders: mental
illness is more visibly associated with women, but
tends to be more harshly derided in men (Chatmon,
2020). This asymmetrical stigma constitutes harms
towards both men and women, increasing the risks
of under-diagnosis or over-diagnosis respectively.
∗Indicates equal contribution.
Figure 1: We investigate masked language models’ bi-
ases at the intersection of gender and mental health.
Using theoretically-motivated prompts about mental
health conditions, we have models fill in the masked to-
ken, then examine the probabilities of generated words
with gender associations.
Since language is central to psychotherapy and
peer support, NLP models have been increasingly
employed on mental health-related tasks (Chancel-
lor and De Choudhury,2020;Sharma et al.,2021,
2022;Zhang and Danescu-Niculescu-Mizil,2020).
Many approaches developed for these purposes rely
on pretrained language models, thus running the
risk of incorporating any pre-learned biases these
models may contain (Straw and Callison-Burch,
2020). However, no prior research has examined
how biases related to mental health stigma are rep-
resented in language models. Understanding if and
how pretrained language models encode mental
health stigma is important for developing fair, re-
sponsible mental health applications. To the best
of our knowledge, our work is the first to opera-
tionalize mental health stigma in NLP research and
aim to understand the intersection between mental
health and gender in language models.
In this work, we propose a framework to inves-
tigate joint encoding of gender bias and mental
arXiv:2210.15144v2 [cs.CL] 11 Apr 2023
health stigma in masked language models (MLMs),
which have become widely used in downstream
applications (Devlin et al.,2019;Liu et al.,2019).
Our framework uses questionnaires developed
in psychology research to curate prompts about
mental health conditions. Then, with several se-
lected language models, we mask out parts of
these prompts and examine the model’s tendency
to generate explicitly gendered words, including
pronouns, nouns, first names, and noun phrases.
1
In order to disentangle general gender biases from
gender biases tied to mental health stigma, we com-
pare these results with prompts describing health
conditions that are not related to mental health.
Additionally, to understand the effects of domain-
specific training data, we investigate both general-
purpose MLMs and MLMs pretrained on mental
health corpora. We aim to answer the two research
questions below.
RQ1: Do MLMs associate mental health con-
ditions with a particular gender?
To answer
RQ1, we curate three sets of prompts that reflect
three healthcare-seeking phases: diagnosis, inten-
tion, and action, based on the widely-cited Health
Action Process Approach (Schwarzer et al.,2011).
We prompt the models to generate the subjects of
sentences that indicate someone is (1) diagnosed
with a mental health condition, (2) intending to
seek help or treatment for a mental health condi-
tion, and (3) taking action to get treatment for a
mental health condition. We find that models asso-
ciate mental health conditions more strongly with
women than with men, and that this disparity is ex-
acerbated with sentences indicating intention and
action to seek treatment. However, MLMs pre-
trained on mental health corpora reduce this gender
disparity and promote gender-neutral subjects.
RQ2: How do MLMs’ embedded preconcep-
tions of stereotypical attributes in people with
mental health conditions differ across genders?
To answer RQ2, we create a set of prompts that de-
scribe stereotypical views of someone with a men-
tal health condition by rephrasing questions from
the Attribution Questionnaire (AQ-27), which is
widely used to evaluate mental health stigma in
psychology research (Corrigan et al.,2003). Then,
using a recursive heuristic, we prompt the mod-
els to generate gendered phrases and compare the
1
We focus most of our analyses on binary genders (female
and male), due to the lack of gold-standard annotations of
language indicating non-binary and transgender. We discuss
more details of this limitation in § 7.
aggregate probabilities of different genders. We
find that MLMs pretrained on mental health cor-
pora associate stereotypes like anger, blame, and
pity more strongly with women than men, while
associating avoidance and lack of help with men.
Our empirical results from these two research
questions demonstrate that models do perpetu-
ate harmful patterns of overlooking men’s mental
health and capture social stereotypes of men be-
ing less likely to receive care for mental illnesses.
However, different models reduce stigma in some
ways and increase it in other ways, which has sig-
nificant implications for the use of NLP in men-
tal health as well as in healthcare in general. In
showing the complex nuances of models’ gendered
mental health stigma, we demonstrate that context
and overlapping dimensions of identity are impor-
tant considerations when assessing computational
models’ social biases and applying these models in
downstream applications.2
2 Background and Related Work
Mental health stigma and gender.
Mental health
stigma can be defined as the negative perceptions of
individuals based on their mental health status (Cor-
rigan and Watson,2002). This definition is implic-
itly composed of two pieces: assumptions about
who may have mental health conditions in the first
place, and assumptions about what such people
are like in terms of characteristics and personal-
ity. Thus, our study at the intersection of gender
bias and mental health stigma is twofold: whether
models associate mental health conditions with a
particular gender, and what presuppositions these
models have towards different genders with mental
illness.
Multiple psychology studies have reported that
mental health stigma manifests differently for dif-
ferent genders (Sickel et al.,2014;Chatmon,2020).
Regarding the first aspect of stigma, mental ill-
ness is consistently more associated with women
than men. The World Health Organization (WHO)
reports a greater number of mental health diag-
noses in women than in men (WHO,2021), but the
fewer diagnoses in men does not indicate that men
struggle less with mental health. Rather, men are
less likely to seek help and are significantly under-
diagnosed, and stigma has been cited as a leading
barrier to their care (Chatmon,2020).
2
Code and data are publicly available at
https://github.
com/LucilleN/Gendered-MH-Stigma-in-Masked-LMs.
Regarding the second aspect of stigma, prior
work in psychology has developed ways to evalu-
ate specific stereotypes towards individuals with
mental illness. Specifically, the widely used attri-
bution model developed by Corrigan et al. (2003)
defines nine dimensions of stigma
3
about people
with mental illness: blame, anger, pity, help, dan-
gerousness, fear, avoidance, segregation, and coer-
cion. The model uses a questionnaire (AQ-27) to
evaluate the respondent’s stereotypical perceptions
towards people with mental health conditions (Cor-
rigan et al.,2003). To the best of our knowledge,
no prior work has examined how these stereotypes
4
differ towards people with mental health conditions
from different gender groups.
Bias research in NLP.
There is a large body
of prior work on bias in NLP models, particularly
focusing on gender, race, and disability (Garrido-
Muñoz et al.,2021;Blodgett et al.,2020;Liang
et al.,2021). Most of these works study bias in
a single dimension as intersectionality is difficult
to operationalize (Field et al.,2021), though a
few have investigated intersections like gender and
race (Tan and Celis,2019;Davidson et al.,2019).
Our methodology follows prior works that used
contrastive sentence pairs to identify bias (Nan-
gia et al.,2020;Nadeem et al.,2020;Zhao et al.,
2018;Rudinger et al.,2018), but unlike existing
research, we draw our prompts and definitions of
stigma directly from psychology studies (Corrigan
et al.,2003;Schwarzer et al.,2011).
Mental health related bias in NLP.
There has
been little work examining mental health bias in
existing models. One relevant work evaluated
mental health bias in two commonly used word
embeddings, GloVe and Word2Vec (Straw and
Callison-Burch,2020). Our project expands upon
this work as we focus on more recent MLMs, in-
cluding general-purpose MLM RoBERTa, as well
as MLMs pretrained on health and mental health
corpora, MentalRoBERTa (Ji et al.,2021) and Clin-
icalLongformer (Li et al.,2022). Another line of
work studied demographic-related biases in mod-
els and datasets used for identifying depression in
3
We use stigma in this paper to refer to public stigma,
which can be more often reflected in language than other types
of stigma: self stigma and label avoidance.
4
Dimensions of stigma refers to the nine dimensions of
public stigma of mental health, stereotypes towards people
with mental health conditions refers to specific stereotypical
perceptions. For example, “dangerousness” is a dimension of
stigma and “people with schizophrenia are dangerous” is a
stereotype.
social media texts (Aguirre et al.,2021;Aguirre
and Dredze,2021;Sherman et al.,2021). These
works focus on extrinsic biases – biases that surface
in downstream applications, such as poor perfor-
mance for particular demographics. Our paper dif-
fers in that we focus on intrinsic bias in MLMs – bi-
ases captured within a model’s parameters – which
can lead to downstream extrinsic biases when such
models are applied in the real world.
3 Methodology
We develop a framework grounded in social psy-
chology literature to measure MLMs’ gendered
mental health biases. Our core methodology
centers around (1) curating mental-health-related
prompts and (2) comparing the gender associations
of tokens generated by the MLMs.
5
In this section,
we discuss methods for the two research questions
introduced in § 2.
3.1 RQ1: General Gender Associations with
Mental Health Status
RQ1 explores whether models associate mental ill-
ness more with a particular gender. To explore
this, we conduct experiments in which we mask
out the subjects
6
in the sentences, then evaluate the
model’s likelihood of filling in the masked subjects
with male, female, or gender-unspecified words,
which include pronouns, nouns, and names. The
overarching idea is that if the model is consistently
more likely to predict a female subject, this would
indicate that the model might be encoding preexist-
ing societal presuppositions that women are more
likely to have a mental health condition. We an-
alyze these likelihoods quantitatively to identify
statistically significant patterns in the model’s gen-
der choices.
Prompt Curation.
We manually construct three
sets of simple prompts that reflect different stages
of seeking healthcare. These stages are grounded
in the Health Action Process Approach (HAPA)
(Schwarzer et al.,2011), a psychology theory that
models how individuals’ health behaviors change.
We develop prompt templates in three different
stages to explore stigma at different parts of the
5
We choose to use mask-filling, as opposed to generating
free text or dialogue responses about mental health, because
mask-filling provides a more controlled framework: there are
a finite set of options to define the mask in a sentence, which
makes it easier to analyze and interpret the results.
6
"Subject" refers to the person being described, which may
or may not be the grammatical subject of the sentence.
process, differentiating being diagnosed from in-
tending to seek care and from actually taking ac-
tion to receive care. For each prompt template,
we create 11 sentences by replacing “[diagnosis]”
with one of the top-11 mental health (MH) or non-
mental-health-related (non-MH) diagnoses (more
details in § 3.3). Example templates and their corre-
sponding health action phases include:
•
Diagnosis:
“<mask> has [diagnosis]”
•
Intention: “<mask> is
looking for a therapist for [diagnosis]”
•
Action:
“<mask> takes medication for [diagnosis]” The full
list of prompts can be found in Appendix A.
Mask Values.
For each prompt, we identify
female, male, and unspecified-gender words in
the model’s mask generations and aggregate their
probabilities (see footnote 1). Most prior work
has primarily considered pronouns as represen-
tations of gender (Rudinger et al.,2018;Zhao
et al.,2018). However, nouns and names are
also common in mental health contexts, such as
online health forums and therapy transcripts. In
fact, some names and nouns frequently appear in
the top generations of masked tokens. Thus, we
look for: (1) Binary-gendered pronouns (e.g., “He”
and “She”). (2) Explicitly gendered nouns (e.g.,
“Father” and “Mother”). We draw this list of 66
nouns from Field and Tsvetkov (2020). (3) Gender-
associated first names (e.g., “David” and “Mary”).
We identify the top 1,000 most common, unam-
biguous male and female first names in Field et al.
(2022)’s Wikipedia data and consider any non-
repeated names in these lists to be gendered. Any
generations that do not fall into the above cat-
egories are considered unspecified-gender (e.g.,
words like “they” and “friend”). For each prompt,
we sum the probabilities of all female, male, and
unspecified-gender words with probabilities higher
than 0.01.
3.2 RQ2: Gender Associations with
Dimensions of Mental Health Stigma
RQ1 focuses on gender association in general, but
does not explore specific, more nuanced stereo-
types or dimensions of stigma, which we examine
in RQ2. RQ2 investigates what personal attributes
language models associate with men and women
with mental illnesses. At a high level, our method-
ology for RQ2 is similar to RQ1: we once again
mask out words that MLMs will fill with a female,
male, or unspecified-gender word, but we use a
different set of prompts and propose a new method
for generating noun phrases.
Prompt Curation.
The Attribution Question-
naire (AQ-27) is widely used in psychology re-
search to evaluate an individual’s level of stigma
towards people with mental illness (Corrigan et al.,
2003). The language used in the questionnaire has
been subsequently validated by years of psychol-
ogy research, so we use these questions as the basis
for creating RQ2’s prompts. AQ-27 begins with a
vignette about a hypothetical man named Harry
who has schizophrenia, followed by 27 Likert-
scale questions to probe the respondent about their
views towards Harry. These 27 questions corre-
spond to 9 dimensions of mental health stigma,
with 3 questions per dimension. Since we want
to curate prompts that do not indicate explicit gen-
der, we rephrase the questions into descriptive sen-
tences. Each sentence describes a stereotypical
view towards a masked-out subject who has a men-
tal health condition. Similarly to RQ1, we use
the same set of 11 most common mental health
diagnoses to create 11 versions of each of these
sentences. Examples of this set of prompts include:
•
“I would feel aggravated by a <mask> who has
[diagnosis].”
•
“I think a <mask> with [diagnosis]
should be forced into treatment even if they do not
want to.”
•
“I feel much sympathy for a <mask>
with [diagnosis].” The full set of prompts is in
Appendix B.
Recursive Masking for Gendered Phrase
Generation.
Some prompts in this set describe
very specific situations, and the probabilities of
generating a single-token gendered subject are rel-
atively low. To reduce the sparsity of generated
gendered subjects, we design a recursive procedure
that enables generating multi-token noun phrases
as follows. First, we pass the model an initial
prompt: e.g. “I feel aggravated by a <mask>
with schizophrenia.” Then, if the model gener-
ates an unspecified-gender subject (e.g. friend), we
prompt the model to generate a linguistic modifier
by adding a mask token directly before the token
generated in step 1: e.g., “I feel aggravated by a
<mask> friend with schizophrenia.”7
7
We repeat step 2 a predefined number of times (
n= 3
),
though
n
can be adjusted to create phrases of different lengths.
Since we mask out the subjects in the prompts, the final gen-
erated tokens are almost always well-formed noun phrases.
At each recursive step, we consider the top 10 generations.
We stop after
n= 3
steps, as generations afterwards have
low probabilities and do not contribute significantly to the
aggregate probabilities.
摘要：
展开>>
收起<<
GenderedMentalHealthStigmainMaskedLanguageModelsInnaWanyinLin1LucilleNjoo1AnjalieField2AshishSharma1KatharinaReinecke1TimAlthoff1YuliaTsvetkov11PaulG.AllenSchoolofComputerScience&Engineering,UniversityofWashington2StanfordUniversity{ilin,lnjoo}@cs.washington.eduAbstractMentalhealthstigmapreventsma...
声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！
相关推荐
-
Michael Moorcock - Elric 6 - StormbringerVIP免费
2024-12-08 6 -
Michael Crichton - PreyVIP免费
2024-12-08 11 -
Mercedes Lackey - WintermoonVIP免费
2024-12-08 7 -
Mercedes Lackey - SE 1- Born To RunVIP免费
2024-12-08 8 -
Mercedes Lackey - Heralds of Valdemar 1 - Arrows Of The QueeVIP免费
2024-12-08 10 -
Melville, Herman - TypeeVIP免费
2024-12-08 15 -
MaryJanice Davidson - [Betsy 5] - Undead and Unpopular (v1.0)VIP免费
2024-12-08 18 -
Marion Zimmer Bradley - Darkover - The Heirs of HammerfellVIP免费
2024-12-08 19 -
MacDonnell, J E - 096 - Execute!VIP免费
2024-12-08 15 -
Lovecraft, H P - The Dream Quest Of Unknown KadadthVIP免费
2024-12-08 21
分类：图书资源
价格：10玖币
属性：20 页
大小：842.65KB
格式：PDF
时间：2025-05-06
作者详情
-
Adversarial and Safely Scaled Question Generation Sreehari Sankar Zhihang Dong October 19 202210 玖币0人下载
-
Adiabatic-impulse approximation in non-Hermitian Landau-Zener Model Xianqi Tong1Gao Xianlong2and Su-peng Kou1y 1Department of Physics Beijing Normal University Beijing 100000 Peoples Republic of China10 玖币0人下载
相关内容
-
2015年6月英语四级真题答案及解析（卷二）
分类：外语学习
时间：2025-05-02
标签：无
格式：PDF
价格：5.8 玖币
-
2015年6月英语四级真题答案及解析（卷三）
分类：外语学习
时间：2025-05-02
标签：无
格式：PDF
价格：5.8 玖币
-
2016年12月六级（第二套）真题
分类：外语学习
时间：2025-05-02
标签：无
格式：PDF
价格：5.8 玖币
-
2016年12月六级（第三套）真题
分类：外语学习
时间：2025-05-02
标签：无
格式：PDF
价格：5.8 玖币
-
2016年12月六级（第一套）真题
分类：外语学习
时间：2025-05-02
标签：无
格式：PDF
价格：5.8 玖币
渝公网安备50010702506394
