Content uploaded by Sidney GJ Wong
Author content
All content in this area was uploaded by Sidney GJ Wong on Jul 01, 2024
Content may be subject to copyright.
Sociocultural Considerations in Monitoring Anti-LGBTQ+ Content on
Social Media
Sidney G.-J. Wong1,2,3
1University of Canterbury, New Zealand
2Geospatial Research Institute, New Zealand
3New Zealand Institute of Language, Brain and Behaviour, New Zealand
{sidney.wong}@pg.canterbury.ac.nz
Abstract
The purpose of this paper is to ascertain the in-
fluence of sociocultural factors (i.e., social, cul-
tural, and political) in the development of hate
speech detection systems. We set out to investi-
gate the suitability of using open-source train-
ing data to monitor levels of anti-LGBTQ+ con-
tent on social media across different national-
varieties of English. Our findings suggests the
social and cultural alignment of open-source
hate speech data sets influences the predicted
outputs. Furthermore, the keyword-search ap-
proach of anti-LGBTQ+ slurs in the develop-
ment of open-source training data encourages
detection models to overfit on slurs; therefore,
anti-LGBTQ+ content may go undetected. We
recommend combining empirical outputs with
qualitative insights to ensure these systems are
fit for purpose.
Content Warning: This paper contains unobfus-
cated examples of slurs, hate speech, and offensive
language with reference to homophobia and trans-
phobia which may cause distress.
1 Introduction
The proliferation of hate speech on social media
platforms continues to negatively impact LGBTQ+
communities (Stefania and Buf,2021). As a conse-
quence of anti-LGBTQ+ hate speech, these already
minoritised and marginalised communities may ex-
perience digital exclusion and barriers to access
in the form of the digital divide (Norris,2001).
There have been considerable developments within
the field of Natural Language Processing (NLP)
in response to this social issue (Sánchez-Sánchez
et al.,2024), with most of the methodological ad-
vancements in this area being made in the last three
decades (Tontodimamma et al.,2021).
While much of hate speech research has focused
on documentation and detection, there has been
little attention on how these approaches can be ap-
plied across different social, political, or linguistic
contexts (Locatelli et al.,2023). Just as the appro-
priateness of swear words is highly contextually
variable depending on language and culture (Jay
and Janschewitz,2008), hate speech in the form
of anti-LGBTQ+ hate speech is often predicated
by social, cultural, and political attitudes towards
diverse gender and sexualities. With minimal litera-
ture beyond just a system development context, we
set out to investigate the suitability of implement-
ing open-source anti-LGBTQ+ hate speech system
on real-world sources of social media data.
This paper makes two contributions: firstly, we
show the predicted outputs from classification mod-
els can be transformed into various time series
data sets to monitor the rate and volume of anti-
LGBTQ+ hate speech on social media. Secondly,
we argue that social, cultural, and linguistic bias
introduced during the data collection phase has an
impact on the suitability of these approaches.
1.1 Related Work
Hate speech detection is often treated as a text clas-
sification task, whereby existing data can be used
to train machine learning models to predict the
attributes of unknown data (Jahan and Oussalah,
2023). The main focus of these systems are racism,
sexism and gender discrimination, and violent rad-
icalism (Sánchez-Sánchez et al.,2024). Both the
production and deployment of hate speech detec-
tion systems are methodologically similar produced
under the following pipeline (Kowsari et al.,2019):
a)
Data Set Collection and Preparation: involves
collecting either real-world or synthetic in-
stances of hate speech in a language condition
(i.e., keyword search). This phase may in-
volve or manual annotation from experts of
crowd-sourced annotators.
b)
Feature Engineering: involves manipulating
and transforming instances of hate speech.
arXiv:submit/5702373 [cs.CL] 1 Jul 2024
This may involve anonymisation or confiden-
tialisation depending on the privacy and data
use rules for each social media platform.
c)
Model Training: involves developing a hate
speech detection system with machine learn-
ing algorithms. This may involve statistical
language models or incorporating transformer-
based large language models.
d)
Model Evaluation: involves producing model
performance metrics to determine the statisti-
cal validity of the system. This may involve
making predictions on unseen or test data.
Despite their straightforward workflow, these
systems pose a number of ethical challenges and
risks to the vulnerable communities (Vidgen and
Derczynski,2020). Cultural biases and harms can
be introduced at each stage of the data set produc-
tion process (Sap et al.,2019). Some of this can
be attributed to poorly designed systems which are
not fit for purpose (Vidgen and Derczynski,2020).
For example, racial bias was identified in one open-
source hate speech detection system developed by
Davidson et al. (2017) which resulted in samples
of written African American English being mis-
classified as instances of hate speech and offensive
language (Davidson et al.,2019).
The presence of racial bias can be attributed to
the decisions made during the Data Set Collec-
tion and Preparation phase during system devel-
opment. Davidson et al. (2017) took a keyword
search approach (i.e., slurs and profanities) to iden-
tify instances of hate speech and offensive language.
These samples were then used in the development
of the detection system. Although slurs and pro-
fanities are good evidence of anti-social behaviour,
the same words can also be re-appropriated or re-
claimed by target communities (Popa-Wyatt,2020).
Classifications algorithms are unable to account for
implicit world knowledge.
Similarly, simple machine learning algorithms
cannot account for linguistic variation which is
another form of implicit world knowledge. Of in-
terest to our current investigation, Wong (2023a)
applied the same system developed by Davidson
et al. (2017) on samples of tweets/posts originat-
ing in New Zealand. The system erroneously
classified tweets/posts with words such as bugger,
digger, and stagger as instances of hate speech.
An unintended consequence of these misclassified
tweets/posts is that rural areas exhibited higher
rates of hate speech and offensive language when
compared to the national mean.
However, not all forms of biases stem from de-
cisions made during system development. Recent
innovations in transformer-based language models,
such as BERT (Devlin et al.,2019), have introduced
new ethical challenges as the presence of gender,
race, and other forms of bias have been observed
in the word embeddings of large language models
(Tan and Celis,2019). This means there is potential
for bias even in the later stages of system develop-
ment during the Model Training phase.
While we grow increasingly aware of the im-
pacts from these limitations (Alonso Alemany et al.,
2023), the number of hate speech detection data
sets and systems continue to increase (Tontodi-
mamma et al.,2021). A systematic review of hate
speech literature has identified over 69 training
data sets to detect hate speech on online and social
media for 21 different language conditions (Jahan
and Oussalah,2023). Seemingly, the solution to
addressing social, cultural, and political discrep-
ancies within hate speech detection is to develop
more systems in different languages.
There remains little interest from NLP re-
searchers to consider the issue of hate speech de-
tection from a social impact lens (Hovy and Spruit,
2016). The primary concerns in this research area
are largely methodological. For example, improv-
ing model performance of detection systems result-
ing from noisy training data (Arango et al.,2022).
Laaksonen et al. (2020) critiqued the datafication
of hate speech detection which in turn has become
an unnecessary distraction for NLP researchers in
combating this social issue.
In fact, the appetite in applying NLP approaches
for social good has decreased over time (Fortuna
et al.,2021). Some researchers are beginning to
question whether the efforts put towards the de-
velopment and production of hate speech detec-
tion systems is the ideal solution for this social
issue (Parker and Ruths,2023). In sidelining these
pressing issues in hate speech detection research,
we may unintentionally perpetuate existing preju-
dices against marginalised and minoritised groups
these systems were meant to support (Buhmann
and Fieseler,2021).
In light of these ethical and methodological chal-
lenges in hate speech detection (Das et al.,2023),
we are starting to see how sociolinguistic infor-
mation can be used to fine tune and improve the
social and cultural performance of hate speech de-
Hostility Direct Indirect Total
Abusive 20 45 65
Disrespectful 5 56 61
Fearful 5 47 52
Hateful 36 106 142
Normal 13 71 84
Offensive 65 308 373
Total 144 633 777
Table 1: The distribution of English posts/tweets and
the level of hostility by directness targeting sexual ori-
entation in Ousidhoum et al. (2019). Note that all totals
are total responses.
tection (Wong et al.,2023;Wong and Durward,
2024) using well-attested methods such as domain
adaptation (Liu et al.,2019). NLP researchers may
still play an invaluable role in combating online
hate speech by incorporating sociocultural consid-
erations in the development and deployment of hate
speech detection systems.
2 Methodology
As discussed in Section 1.1, hate speech detection
research needs to undergo a paradigmatic shift in
order to truly enable positive social impact, social
good, and social benefit potential. The main pur-
pose of this paper is to ascertain the influence of
sociocultural factors (i.e., social, cultural, and po-
litical) in the development of hate speech detection
systems. Our research questions are as follows:
RQ1
Can we use open-source hate speech training
data to monitor anti-LGBTQ+ hate speech in
real world instances of social media? and;
RQ2
How do the social, cultural, and linguistic con-
texts of open-source training data impact on
the suitability of anti-LGBTQ+ hate speech
detection?
In order to address RQ1, we compare and con-
trast two anti-LGBTQ+ hate speech detection sys-
tems. We provide an in depth description of the
data sources in Section 2.1 and our system devel-
opment pipeline in Section 2.2. Once we develop
the detection systems, we apply the detection sys-
tems on real-world samples of social media data to
monitor anti-LGBTQ+ hate speech across different
geographic dialects.
We opted for a mixed-methods approach to ad-
dress this emergent area of enquiry. This is because
Class EN G TAM TAM-ENG
HOMO 276 723 465
TRANS 13 233 184
NONE 4,657 3,205 5,385
Total 4,946 4,161 6,034
Table 2: The class distribution of YouTube comments
based on the three-class classification system (homo-
phobic (HO MO), transphobic (TRANS), and non-anti-
LGBTQ+ (NO NE) content) by language condition (En-
glish (EN G), Tamil (TAM), and Tamil-English (TAM-
ENG)) in Chakravarthi et al. (2021).
RQ2 can only be addressed qualitatively as we con-
sider the suitability of the detection systems and
the sociocultural relevance of the predicted outputs.
We will address RQ2 in the discussion (Section 4);
however, we have provided relevant sociolinguistic,
cultural, and political information in Section 2.3 to
contextualise our discussion.
2.1 Data Sources
As part of our investigation, we use two open-
source training data sets to develop our anti-
LGBTQ+ hate speech detection systems in our
investigation: Ousidhoum et al. (2019) (Multi-
lingual and Multi-Aspect Hate Speech Data Set;
MLMA) and Chakravarthi et al. (2021) (LTEDI)
1
.
The MLMA and LTEDI were chosen due to the avail-
ability of data and documentation to understand the
data set collection and annotation process.
The MLMA is a multilingual hate speech data
set for posts/tweets from X (Twitter) for English,
French, and Arabic (Ousidhoum et al.,2019). The
authors took a keyword search approach by retriev-
ing posts/tweets which matched a list of common
slurs, controversial topics, and discourse patterns
typically found in a hate speech. This approach
proved challenging due to the high-rates of code-
switching in the English and French conditions
and Arabic diglossia. The posts/tweets were then
posted on the crowd-sourcing platform, Mechani-
cal Turk, for public annotation.
One of the most well-documented anti-LGBTQ+
training data sets is the English, Tamil, and English-
Tamil anti-LGBTQ+ hate speech data set developed
by Chakravarthi et al. (2021). The data set contains
public comments to LGBTQ+ videos on YouTube.
The comments were manually annotated based on
1
We refer to it as LTEDI with reference to its central role
in the various shared tasks hosted as part of the Language
Technology for Equity, Diversity, and Inclusion
Figure 1: Model comparison of anti-LGBTQ+
hate speech on ten randomised samples of 10,000
posts/tweets per month from India between June 2018
to June 2023 including grouped mean and the upper and
lower confidence intervals.
a three-class (i.e., homophobic, transphobic, and
non-anti-LGBTQ+ hate speech). The training data
was tested with three language models: MURIL
(Khanuja et al.,2021), MBERT (Pires et al.,2019),
and XLM-ROB ERTA (Conneau et al.,2020).
The results show that transformer-based models,
such as BERT, outperformed statistical language
models with minimal fine-tuning. The best per-
forming BERT-based system for English yielded an
averaged
F1
-score of 0.94 (Maimaitituoheti et al.,
2022). This anti-LGBTQ+ training data set has
since expanded to a suite of additional language
conditions such as Spanish (García-Díaz et al.,
2020), Hindi and Malayalam (Kumaresan et al.,
2023), and Telugu, Kannada, Gujarati, Marathi,
and Tulu (Chakravarthi et al.,2024).
We discuss the similarities and differences be-
tween the two data sets in relation sociocultural
considerations regarding the data collection strat-
egy in Section 2.1.1, the annotation strategy in Sec-
tion 2.1.2, and the cultural alignment in Section
2.1.3 derived from available documentation.
2.1.1 Data Collection
The developers of the MLMA took a culturally-
agnostic approach with limited information on the
data collection points; however, evidence of code-
switching between English with Hindi, Spanish,
and French posed a challenge to annotators. The
MLMA took a keyword search approach to filter
X (Twitter) for instances hate speech. The key-
words in relation to anti-LGBTQ+ hate in English
included: dyke,twat, and faggot. This contrasts
LTEDI which took a content search approach of
users reacting to LGBTQ+ content from India.
The high-level of code-switching and script-
switching between English and other Indo-Aryan
and Dravidian languages provides some level of
social, cultural, and linguistic information of the
training data. Both training data sets are compara-
ble in size; however, MLMA is 13.2% larger than
LTEDI by number of observations. The proportion
of anti-LGBTQ+ hate speech in the MLMA is 9.1%
while the proportion of anti-LGBTQ+ hate speech
in the LTEDI is 5.8%.
2.1.2 Annotation Process
Bender and Friedman (2018) proposed including
data statement framework in the hope to mitigate
different forms of social bias by dutifully doc-
umenting the NLP production process. Neither
data sets provided annotator metadata (Bender and
Friedman,2018); therefore, we can only infer
some of the annotator information from available
documentation. Where the MLMA took a crowd-
sourcing approach, the LTEDI data set were an-
notated by members of the LGBTQ+ communi-
ties. Based on the limited details, LTEDI we know
the annotators were English speakers based at the
National University of Ireland Galway. Unsurpris-
ingly, the MLMA at 0.15 is lower than LTEDI at 0.67
based on Krippendorf’s alpha where 1 suggests
perfect reliability while 0 suggests no reliability
beyond chance.
2.1.3 Cultural Alignment
With limited documentation to the data set col-
lection and annotation process beyond the system
description papers, we tentatively determine the
LTEDI is largely in alignment with anti-LGBTQ+
discourse from the South Asian cultural sphere
and the MLMA as culturally-undetermined anti-
LGBTQ+ rhetoric. This creates a useful contrast
which not only compares the efficacy of two train-
ing data sets, but also anti-LGBTQ+ behaviour in
different varieties of World Englishes which are
influenced by their own unique social, cultural, and
linguistic contexts (Kachru,1982). We predict the
data set collection and annotation approaches will
have an impact on the outputs of the automatic
detection systems.
2.2 System Development
The first phase of our investigation involves de-
veloping multiclass classification models to detect
Figure 2: Comparison of anti-LGBTQ+ hate speech detected in 10,000 samples of posts/tweets from inner- and
outer-circle varieties of English between June 2018 to June 2023 including grouped mean and the upper and lower
confidence intervals.
Macro Weighted
Base Retrain Base Retrain
LTEDI 0.78 0.81 0.95 0.96
MLMA 0.83 0.83 0.94 0.94
Table 3: Model evaluation metrics comparing the four
candidate models by average macro
F1
-score and aver-
age weighted F1-score.
anti-LGBTQ+ hate speech in English. We opted for
a transformer-based language modelling approach.
Even though the focus of LTEDI is YouTube, we
can adapt Pretrained Language Models (PLMs) to
specific domains, or register of language, through
pretraining with additional samples of text (Guru-
rangan et al.,2020).
We initially trained two classification models
with minimal feature engineering in order to de-
termine the best approaches to develop our auto-
matic detection systems. We split the training data
into training, development, and test sets with a
train:development:test split of 90:5:5. We used
Multi-Class Classification model from the Simple
Transformers
2
Python package to finetune and train
the multi-class classification model. We trained
each model for 8 iterations. We used AdamW as
the optimiser (Loshchilov and Hutter,2018). Our
baseline PLM is XLM-ROBERTA, which is a cross-
lingual transformer-based language model (Con-
neau et al.,2020).
2.2.1 Feature Engineering
Class imbalance had an effect on our detection sys-
tem. Therefore, we collapsed the multiple classes
from each training data set into a binary classifi-
cation. We also removed the confidentialised user-
2https://simpletransformers.ai/
names and URLs from Ousidhoum et al. (2019),
as we could not mask these high-frequency to-
kens from the classification model. We used
RandomOverSampler from the Imbalanced Learn
3
Python package to upsample the minority classes.
We address the register discrepancy in Chakravarthi
et al. (2021). We retrained XLM-ROB ERTA with
120,000 samples of X (Twitter) language data from
the CGLU (Dunn,2020). The composition of the
language data included 10,000 samples from each
language condition.
2.2.2 Model Evaluation
We present the model evaluation metrics in Table 3.
In Table 3, we compare the model evaluation results
for the four candidate models (LTEDI
B
,LTEDI
R
,
MLMA
B
, and MLMA
R
). The model performance
improved in three of the four candidate models
based on both macro average and weighted average
F1
-score. Surprisingly, there were no differences
between the two approaches for the MLMA mod-
els. With a focus on the anti-LGBTQ+ class, do-
main adaptation improved the
F1
-score from 0.58
to 0.64 for the LTEDI
R
model. The
F1
-score for the
MLMA
R
remains unchanged at 0.69. Based on the
model performance metrics for the four candidate
models, we advanced with the LTEDI
R
and MLMA
R
classification models with domain adaptation and
feature engineering during finetuning. We contin-
ued to apply domain adaptation in both systems
despite not seeing significant improvements in the
MLMA
R
model to maintain consistency between
the two classification models.
3https://imbalanced-learn.org/
Figure 3: Quarterly growth rate of anti-LGBTQ+ hate speech detected with the LTEDI model with number of
posts/tweets by country between June 2018 and June 2023.
2.3 Communities of Interest
Even though the MLMA is supposedly culturally-
agnostic, we have broadly identified the cultural
alignment within the LTEDI based on the data
set collection and annotation process outlined in
Chakravarthi et al. (2021). More specifically, high-
levels of code-switching and script-switching be-
tween English, Hindi, and Tamil in the LTEDI sug-
gests the presence of an Indian English substrate in
the training data. Written English is often treated
as homogeneous language; however, geographic-
dialects represented by national-varieties of En-
glish maintain a constant-level of variation (Dunn
and Wong,2022).
Furthermore, the presence of Indian English on
social media, or English spoken and written in In-
dia introduced as a result of British colonisation
(Hickey,2005), is uncontested (Rajee,2024). In the
three concentric circles model of World Englishes,
Indian English is categorised as an outer-circle va-
riety of English (Kachru et al.,1985). Outer-circle
and inner-circle varieties of English are defined as
national-varieties with British colonial ties. The
distinguishing feature of outer-circle varieties is
that English is not the primary language of social
life and the government sector. These outer-circle
varieties of English often co-exist alongside other
indigenous languages.
In order to test for the influence of social, cul-
tural, and linguistic factors, we retrieved samples of
social media language from outer-circle and inner-
circle varieties of English. Outer-circle varieties of
English as written English originating from Ghana,
India, Kenya, Malaysia, the Philippines, and Pak-
istan. Similarly, inner-circle varieties as written
English originating from Australia, Canada, Ire-
land, New Zealand, the United Kingdom, and the
United States. The data source of our social media
language data comes from a subset of CGLU corpus
which contains georeferenced posts/tweets from X
(Twitter) (Dunn,2020).
For each national-variety of English, we filtered
the data for tweets in English. All posts/tweets
were processed with hyperlinks, emojis, and user
identifying information removed. In addition to the
monthly samples for each country, we re-sampled
monthly tweets from India over ten iterations to
determine the impact of our sampling methodol-
ogy. All posts/tweets were produced between July
2018 to June 2023. Of relevance to our analy-
sis, the countries associated with these national-
varieties all criminalised same-sex sexual activity
as a legacy of the English common law legal system
(with the exception of the Philippines) (Han and
O’Mahoney,2014). All but four of these countries
(Kenya, Ghana, Pakistan, Malaysia) have since
decriminalised same-sex sexual activity. How-
ever, LGBTQ+ rights vary significantly between
countries and LGBTQ+ communities continue to
face discrimination in response to increased anti-
LGBTQ+ legislation in the United States dispro-
portionately affecting transgender people (Canady,
2023).
3 Results
We dedicate the current section to describe the re-
sults of the second phase of our investigation. This
phase involved applying the candidate models to
automatically detect anti-LGBTQ+ hate speech on
real-world instances of social media data in English.
Firstly, we applied both anti-LGBTQ+ hate speech
detection models on the ten randomised monthly
samples of social media language data from In-
dia using the same sampling methodology for other
national-varieties of English. The results are shown
in Figure 1. As expected, the LTEDI
R
model pre-
dicted higher rates of anti-LGBTQ+ hate speech;
however, what was unexpected were the low num-
ber of predictions from the MLMA
R
model. The
narrow confidence intervals suggest little instability
between the different samples and the predictions
remained constant across samples.
After validating our sampling methodology by
visually inspecting the ten randomised monthly
samples from India, we applied both models on
random samples of inner- and outer-circle varieties
of English. We compared the results of the detec-
tion models as visualised in Figure 2. These were
consistent with our initial results. The rate of anti-
LGBTQ+ hate speech remained constant according
to the MLMA
R
model, while anti-LGBTQ+ hate
speech has increased over time based on a visual
inspection of the results. Of interest to our investi-
gation, the MLMA
R
model identified a higher pro-
portion of anti-LGBTQ+ hate speech in inner-circle
varieties of English. We saw an inverse relationship
with the LTEDI
R
where we see a higher proportion
of anti-LGBTQ+ hate speech in outer-circle vari-
eties of English. The wide confidence intervals of
the LTEDI
R
suggests greater between-variety insta-
bility in outer-circle varieties of English.
We calculated the quarterly growth rates for
(a) Training Data
(b) Predicted Outputs
Figure 4: MLMA Wordcloud.
each variety of English for the predictions from
the LTEDI
R
. We included the total number of pre-
dicted posts/tweets in our visualisation as shown in
Figure 3. The growth rates allowed us to determine
the growth rate for each variety of English inde-
pendently. The results suggest the growth rate of
predicted anti-LGBTQ+ hate speech has remained
stable over time.
4 Discussion
The results from our study raises some interest-
ing questions on the efficacy of these systems on
real-world instances of social media data. With re-
gards to the first research question, our transformer-
based multiclass classification model enabled us
to detect instances of anti-LGBTQ+ hate speech
from samples of georeferenced posts/tweets from
X (Twitter). We were able to manipulate the pre-
dicted outputs into different forms of time series
as shown in Figures 2and 3. The level of anti-
LGBTQ+ hate speech has maintained a constant
rate of growth despite decreasing usership on the
social media platform since the acquisition of X
(Twitter) by Elon Musk in 2022. The results sug-
gest that anti-LGBTQ+ hate speech on X (Twitter)
is indeed increasing in both rate and volume over
time (Hattotuwa et al.,2023).
When we compare the predicted results between
the MLMA
R
and the LTEDI models, we can see
significant differences between the two models.
This is particularly obvious when we compare
(a) Training Data
(b) Predicted Outputs
Figure 5: LTEDI Wordcloud.
the predicted outputs in Figures 1and 2, where
the LTEDI
R
model on average predicted 50 times
more instances of anti-LGBTQ+ hate speech than
the MLMA
R
. These was unexpected as the model
evaluation metrics during model development sug-
gested the MLMA
R
model performed marginally
better than the LTEDI
R
. Considering both the sam-
pling methodology and the model development ap-
proaches were held constant between the models,
we propose the differences we see in the predicted
outputs is a result of the open-source training data.
One challenge of applying multiclass classifica-
tion models on unknown data is that there is no sim-
ple method to validate the results. This is because
we do not have access to labelled training, devel-
opment, and test sets to evaluate the model perfor-
mance. We are therefore reliant on qualitative meth-
ods to validate the performance of our detection
models. Figure 4is a visual representation of the
word-token frequencies between the open-source
training data (a) and the predicted anti-LGBTQ+
hate speech (b) from the samples of posts/tweets.
The most prominent word-token in the training data
is faggot followed by dyke. This is not unexpected
as these word-tokens (including twat) were used to
identify instances of anti-LGBTQ+ hate speech on
X (Twitter). Counterintuitively, we did not see a
similar distribution in the predicted outputs.
With reference to Figure 4, the word-tokens with
the highest frequency in the predicted output were
not faggot or dyke, but sleep and gay. When we
Variety dyke faggot twat
GH 8 2 6
IN 5 - 6
KE 1 4 7
MY 3 4 14
PH 8 4 8
PK 3 6 6
gay
353
226
295
500
701
478
Table 4: Frequency of LGBTQ+ related slurs for outer
circle varieties of English.
filtered for the keyword search terms in the sam-
ples, we found few instances across the varieties
of English as shown in Tables 4and 5. This is
unexpected as the keyword search terms are highly
prevalent in inner-circle varieties of spoken En-
glish (such as the United Kingdom and Ireland)
(Love,2021). This is supported by the higher word-
token frequencies in inner-circle varieties of En-
glish as shown in Tables 4and 5. We attribute the
infrequent occurrence of LGBTQ+ slurs in direct
response to X (Twitter) rules which discourages
hateful conduct on the platform.
Our analysis of the MLMA
R
model suggests a re-
lationship between the training data and the result-
ing detection model. Incidentally, we also observe
this bias towards inner-circle varieties of English
in Figure 2where the MLMA
R
is more inclined to
identify more anti-LGBTQ+ hate speech in inner-
circle than outer-circle varieties of English. This
leads our discussion to the second research ques-
tion where we determine how the social, cultural,
and linguistic context impacts the efficacy of anti-
LGBTQ+ hate speech detection. Although anti-
LGBTQ+ discourse is consistent across languages
(Locatelli et al.,2023), slurs and swearwords are
not (Jay and Janschewitz,2008). This form of cul-
tural bias toward inner-circle varieties of English
(or oversight of outer-circle varieties) introduced
during the data collection process, raises questions
on the suitability of the MLMA
R
model in monitor-
ing anti-LGBTQ+ hate speech.
As we determined the LTEDI
R
model to be more
culturally aligned with the South Asian context, we
initially predicted the LTEDI model would be more
appropriate for South Asian contexts. However,
the results suggest the LTEDI
R
model as more fit
for purpose in contrast to the MLMA
R
model. Not
only do we observe high-congruency between the
LTEDI
R
model output and the outer-circle varieties
of English as shown in Figure 2, the word-token
Variety dyke faggot twat
AU 5 12 48
CA 16 13 19
IE 15 16 62
NZ 6 14 53
UK 23 9 148
US 19 11 13
gay
635
623
659
627
679
875
Table 5: Frequency of LGBTQ+ related slurs for inner
circle varieties of English.
frequencies between the training data (a) and the
predicted outputs (b) in the LTEDI appear to have a
similar distribution as shown in Figure 5.
Curiously, both the training data and predicted
output lack slurs. Instead, we see word-tokens asso-
ciated with community (e.g., people) and religion
(e.g., bible,god, and Adam possibly in reference
to the Abrahamic creation myth of Adam and Eve).
This is unsurprising as anti-LGBTQ+ legislation
is often rooted in puritanical beliefs on morality
(Han and O’Mahoney,2014). With reference to
Figure 3, we observed a possible link between the
increased growth rate with nationwide response to
the Covid-19 pandemic. Once again this raises a
question on the validity of the predicted outputs
and whether the posts/tweets are anti-LGBTQ+ or
religious/spiritual in nature (or indeed, both).
5 Conclusion
The findings from this current paper raises a num-
ber challenges in applying hate speech detection in
a real-world context. Even within national-varieties
of English, we observed the impacts of social,
cultural, and linguistic factors. For example, the
LTEDI
R
which was culturally aligned with Indian
English was more sensitive to outer circle varieties
of English, while the MLMA
R
model was slightly
more sensitive to inner circle varieties of English.
We conclude that monitoring anti-LGBTQ+ hate
speech with open-source training data is not prob-
lematic in itself; however, we must interpret these
empirical outputs with qualitative insights to ensure
these systems are fit for purpose.
Ethics Statement
The purpose of this paper is to investigate the suit-
ability of using open-source training data to de-
velop a multiclass classification model to monitor
and forecast levels of anti-LGBTQ+ hate speech
on social media across different geographic dialect
contexts in English. This study contributes to the ef-
forts in mitigating harmful hate speech experienced
by LGBTQ+ communities. In our investigation, we
combine methods from NLP, sociolinguistics, and
discourse analysis to evaluate the effectiveness of
anti-LGBTQ+ hate speech detection.
We recognise the importance of advocate and
activist-led research in particular by members of
under-represented and minoritised communities
(Hale,2008). The lead author acknowledges their
positionality as an active advocate and a member of
the LGBTQ+ community (Wong,2023b). The lead
author is familiar with anti-LGBTQ+ discourse
both in online and offline spaces and its harmful
effects on members of the LGBTQ+ communities.
As discussed in Section 5, we support the cri-
tique of Parker and Ruths (2023) for NLP re-
searchers to reflect on the efficacy and suitability
of hate speech detection models. The development
of hate speech data sets impose a ‘diversity tax’ on
already marginalised LGBTQ+ communities. Orig-
inally coined by Padilla (1994), this refers to the un-
intentional burden placed on marginalised peoples
to address inequities, exclusion, and inaccessibility
particularly in a research context. NLP researchers
need to work alongside key-stakeholders (e.g., af-
fected communities, advocates, and activists) as
well as social media platforms, non-profit organi-
sations, and government entities to determine the
solutions of this social issue.
The inclusion of unobfuscated examples of
slurs, hate speech, and offensive language to-
wards LGBTQ+ communities is a deliberate at-
tempt to initiate the process of reclaiming and re-
appropriating some anti-LGBTQ+ slurs in NLP
research. Currently, there are limited best practice
guidelines on the obfuscation of profanities in NLP
research (Nozza and Hovy,2023). Worthen (2020)
theorised that anti-LGBTQ+ slurs are used to stig-
matise violations of social norms. Re-appropriating
these stigmatising labels can enhance what were
once devalued social identities (Galinsky et al.,
2003). This process of ‘cleaning’ and ‘detoxifying’
slurs is also a process of resistance and to reclaim
power and control (Popa-Wyatt,2020).
We argue that within context of social media
research giving unwarranted attention to slurs ig-
nores the root of this social issue: hate speech
expresses hate (Marques,2023). Many social me-
dia platforms have already put in place procedures
to censor sensitive word-tokens; however, social
media users continue to adopt innovative linguistic
strategies such as voldermorting (van der Nagel,
2018) and Algospeak (Steen et al.,2023) to con-
travene well-meaning moderation and censorship
algorithms. Our results suggest hate speech train-
ing data sets do not identify the full breadth of
hateful content on social media.
This paper does not include human or animal
participants. Furthermore, we abide by the data
sharing rules of X (Twitter) and posts/tweets with
identifiable personal details will not be shared pub-
licly. The authors have no conflicts of interests to
declare.
Limitations
In this section, we address some of the known limi-
tations of our approach in addition to limitations of
the open-source training data and the social media
data we have used in the current study.
Invisibility of Q+ identities This paper uses the
LGBTQ+ acronym to signify diverse gender and
sexualities who continue to experience forms of
discrimination and stigmatisation (namely Lesbian,
Gay, Bisexual, and Transgender people). While the
Q+ refers to those who are not straight or not cis-
gender (Queer+), we acknowledge the invisibility
of other minorities who are often excluded from
NLP research including intersex and indigenous
expressions of gender, sexualities, and sex charac-
teristics at birth.
Sociocultural bias during data collection De-
spite including more training data, the MLMA iden-
tified significantly fewer instances of anti-LGBTQ+
hate speech than the LTEDI across the national-
varieties of English. With reference to the word-
clouds produced from the training data for MLMA
and LTEDI as shown in Figures 4and 5, there is
a high likelihood the keyword search (on dyke,
twat, and faggot) during the data collection pro-
cess has caused the classification model to over-fit
the training data. Similarly, the religious subtext
in the LTEDI training data reinforces polarising be-
liefs that religion is anti-LGBTQ+. Furthermore,
these detection systems do not account for semantic
bleaching or the reclamation of slurs (Popa-Wyatt,
2020).
Pitfalls of large language models We acknowl-
edge the cultural and linguistic biases introduced
through the PLMs used in our transformer-based
approach. However, we have mitigated some of
these impacts through domain adaptation (Liu et al.,
2019). With reference to Figure 4, we have reason
to believe the transformer-based detection systems
erroneously classified dylan,mike and like with
dyke. A breakdown of the character-trigrams (#DY,
DYK,YKE, and #KE) confirms this belief.
Class imbalance and distribution We were able
to improve the performance of the detection model
during model development by up-sampling the mi-
nority classes. The LTEDI detected a constant pro-
portion of anti-LGBTQ+ hate speech between 5-
10% for all varieties of English which is a sim-
ilar proportion of anti-LGBTQ+ hate speech in
the training data (or 5.8% of the training data).
This raises potential questions on the efficacy of
transformer-based classification models.
Further work We welcome NLP researchers to
address these limitations in their research especially
on increasing the visibility of Q+ communities and
the sociocultural biases shown in open-source train-
ing data sets and large language models.
Acknowledgements
The lead author wants to thank Dr. Benjamin
Adams (University of Canterbury | Te Whare
W
¯
ananga o Waitaha) and Dr. Jonathan Dunn (Uni-
versity of Illinois Urbana-Champaign) for their
feedback on the initial manuscript. The lead author
wants to thank the three anonymous peer reviewers
and the programme chairs for their constructive
feedback. Lastly, the lead author wants to thank
Fulbright New Zealand | Te T
¯
u
¯
apapa M
¯
atauranga
o Aotearoa me Amerika and their partnership with
the Ministry of Business, Innovation, and Em-
ployment | H
¯
ıkina Whakatutuki for their support
through the Fulbright New Zealand Science and
Innovation Graduate Award.
References
Laura Alonso Alemany, Luciana Benotti, Hernán
Maina, Lucía Gonzalez, Lautaro Martínez, Beatriz
Busaniche, Alexia Halvorsen, Amanda Rojo, and
Mariela Rajngewerc. 2023. Bias assessment for ex-
perts in discrimination, not in computer science. In
Proceedings of the First Workshop on Cross-Cultural
Considerations in NLP (C3NLP), pages 91–106,
Dubrovnik, Croatia. Association for Computational
Linguistics.
Aymé Arango, Jorge Pérez, and Barbara Poblete. 2022.
Hate speech detection is not as easy as you may think:
A closer look at model validation (extended version).
Information Systems, 105:101584.
Emily M. Bender and Batya Friedman. 2018. Data
Statements for Natural Language Processing: Toward
Mitigating System Bias and Enabling Better Science.
Transactions of the Association for Computational
Linguistics, 6:587–604.
Alexander Buhmann and Christian Fieseler. 2021. To-
wards a deliberative framework for responsible inno-
vation in artificial intelligence.Technology in Society,
64:101475.
Valerie A. Canady. 2023. Mounting anti-LGBTQ+
bills impact mental health of youths.Mental Health
Weekly, 33(15):1–6.
Bharathi Raja Chakravarthi, Prasanna Kumaresan, Ruba
Priyadharshini, Paul Buitelaar, Asha Hegde, Hosa-
halli Shashirekha, Saranya Rajiakodi, Miguel Án-
gel García, Salud María Jiménez-Zafra, José García-
Díaz, Rafael Valencia-García, Kishore Ponnusamy,
Poorvi Shetty, and Daniel García-Baena. 2024.
Overview of Third Shared Task on Homophobia and
Transphobia Detection in Social Media Comments.
In Proceedings of the Fourth Workshop on Language
Technology for Equality, Diversity, Inclusion, pages
124–132, St. Julian’s, Malta. Association for Compu-
tational Linguistics.
Bharathi Raja Chakravarthi, Ruba Priyadharshini,
Rahul Ponnusamy, Prasanna Kumar Kumaresan,
Kayalvizhi Sampath, Durairaj Thenmozhi, Sathi-
yaraj Thangasamy, Rajendran Nallathambi, and
John Phillip McCrae. 2021. Dataset for Identifi-
cation of Homophobia and Transophobia in Mul-
tilingual YouTube Comments.arXiv preprint.
ArXiv:2109.00227 [cs].
Alexis Conneau, Kartikay Khandelwal, Naman Goyal,
Vishrav Chaudhary, Guillaume Wenzek, Francisco
Guzmán, Edouard Grave, Myle Ott, Luke Zettle-
moyer, and Veselin Stoyanov. 2020. Unsupervised
Cross-lingual Representation Learning at Scale. In
Proceedings of the 58th Annual Meeting of the Asso-
ciation for Computational Linguistics, pages 8440–
8451, Online. Association for Computational Lin-
guistics.
Dipto Das, Shion Guha, and Bryan Semaan. 2023. To-
ward Cultural Bias Evaluation Datasets: The Case
of Bengali Gender, Religious, and National Iden-
tity. In Proceedings of the First Workshop on Cross-
Cultural Considerations in NLP (C3NLP), pages 68–
83, Dubrovnik, Croatia. Association for Computa-
tional Linguistics.
Thomas Davidson, Debasmita Bhattacharya, and Ing-
mar Weber. 2019. Racial Bias in Hate Speech and
Abusive Language Detection Datasets. In Proceed-
ings of the Third Workshop on Abusive Language
Online, pages 25–35, Florence, Italy. Association for
Computational Linguistics.
Thomas Davidson, Dana Warmsley, Michael Macy, and
Ingmar Weber. 2017. Automated Hate Speech Detec-
tion and the Problem of Offensive Language.arXiv
preprint. ArXiv:1703.04009 [cs].
Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
Kristina Toutanova. 2019. BERT: Pre-training of
Deep Bidirectional Transformers for Language Un-
derstanding. In Proceedings of NAACL-HLT, pages
4171–4186.
Jonathan Dunn. 2020. Mapping languages: the Corpus
of Global Language Use.Language Resources and
Evaluation, 54(4):999–1018.
Jonathan Dunn and Sidney Wong. 2022. Stability of
Syntactic Dialect Classification over Space and Time.
In Proceedings of the 29th International Confer-
ence on Computational Linguistics, pages 26–36,
Gyeongju, Republic of Korea. International Com-
mittee on Computational Linguistics.
Paula Fortuna, Laura Pérez-Mayos, Ahmed AbuRa’ed,
Juan Soler-Company, and Leo Wanner. 2021. Car-
tography of Natural Language Processing for Social
Good (NLP4SG): Searching for Definitions, Statis-
tics and White Spots. In Proceedings of the 1st Work-
shop on NLP for Positive Impact, pages 19–26, On-
line. Association for Computational Linguistics.
Adam D Galinsky, Kurt Hugenberg, Carla Groom, and
Galen V Bodenhausen. 2003. The reappropriation of
stigmatizing labels: Implications for social identity.
In Jeffrey Polzer, editor, Identity Issues in Groups,
volume 5 of Research on Managing Groups and
Teams, pages 221–256. Emerald Group Publishing
Limited.
José Antonio García-Díaz, Ángela Almela, Gema
Alcaraz-Mármol, and Rafael Valencia-García. 2020.
UMUCorpusClassifier: Compilation and evaluation
of linguistic corpus for Natural Language Process-
ing tasks.Procesamiento del Lenguaje Natural,
65(0):139–142.
Suchin Gururangan, Ana Marasovi´
c, Swabha
Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey,
and Noah A. Smith. 2020. Don’t Stop Pretraining:
Adapt Language Models to Domains and Tasks.
arXiv preprint. ArXiv:2004.10964 [cs].
Charles R. Hale. 2008. Engaging Contradictions: The-
ory, Politics, and Methods of Activist Scholarship.
In Engaging Contradictions. University of California
Press.
Enze Han and Joseph O’Mahoney. 2014. British
colonialism and the criminalization of homosexu-
ality.Cambridge Review of International Affairs,
27(2):268–288.
Sanjana Hattotuwa, Kate Hannah, and Kayli Tay-
lor. 2023. Transgressive transitions: Transphobia,
community building, bridging, and bonding within
Aotearoa New Zealand’s disinformation ecologies
march-April 2023. Technical report, The Disinfor-
mation Project, New Zealand.
Raymond Hickey, editor. 2005. Legacies of Colonial
English: Studies in Transported Dialects. Studies
in English Language. Cambridge University Press,
Cambridge.
Dirk Hovy and Shannon L. Spruit. 2016. The Social
Impact of Natural Language Processing. In Proceed-
ings of the 54th Annual Meeting of the Association
for Computational Linguistics (Volume 2: Short Pa-
pers), pages 591–598, Berlin, Germany. Association
for Computational Linguistics.
Md Saroar Jahan and Mourad Oussalah. 2023. A sys-
tematic review of hate speech automatic detection
using natural language processing.Neurocomputing,
546:126232.
Timothy Jay and Kristin Janschewitz. 2008. The prag-
matics of swearing.Journal of Politeness Research
Language Behaviour Culture, 4(2):267–288.
Braj B. Kachru. 1982. The Other tongue: English
across cultures. University of Illinois Press, Urbana-
Champaign.
Braj B. Kachru, R. Quirk, and H. G. Widdowson. 1985.
Standards, codification and sociolinguistic realism.
World Englishes. Critical Concepts in Linguistics,
pages 241–270.
Simran Khanuja, Diksha Bansal, Sarvesh Mehtani,
Savya Khosla, Atreyee Dey, Balaji Gopalan,
Dilip Kumar Margam, Pooja Aggarwal, Rajiv Teja
Nagipogu, Shachi Dave, Shruti Gupta, Subhash
Chandra Bose Gali, Vish Subramanian, and Partha
Talukdar. 2021. MuRIL: Multilingual Represen-
tations for Indian Languages.arXiv preprint.
ArXiv:2103.10730 [cs].
Kamran Kowsari, Kiana Jafari Meimandi, Mojtaba Hei-
darysafa, Sanjana Mendu, Laura Barnes, and Donald
Brown. 2019. Text Classification Algorithms: A
Survey.Information, 10(4):150.
Prasanna Kumar Kumaresan, Rahul Ponnusamy, Ruba
Priyadharshini, Paul Buitelaar, and Bharathi Raja
Chakravarthi. 2023. Homophobia and transphobia
detection for low-resourced languages in social me-
dia comments.Natural Language Processing Jour-
nal, 5:100041.
Salla-Maaria Laaksonen, Jesse Haapoja, Teemu Kin-
nunen, Matti Nelimarkka, and Reeta Pöyhtäri. 2020.
The Datafication of Hate: Expectations and Chal-
lenges in Automated Hate Speech Monitoring.Fron-
tiers in Big Data, 3.
Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man-
dar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
Luke Zettlemoyer, and Veselin Stoyanov. 2019.
RoBERTa: A Robustly Optimized BERT Pretrain-
ing Approach.arXiv preprint. ArXiv:1907.11692
[cs].
Davide Locatelli, Greta Damo, and Debora Nozza. 2023.
A Cross-Lingual Study of Homotransphobia on Twit-
ter. In Proceedings of the First Workshop on Cross-
Cultural Considerations in NLP (C3NLP), pages 16–
24, Dubrovnik, Croatia. Association for Computa-
tional Linguistics.
Ilya Loshchilov and Frank Hutter. 2018. Decoupled
Weight Decay Regularization. In International Con-
ference on Learning Representations.
Robbie Love. 2021. Swearing in informal spoken En-
glish: 1990s–2010s.Text & Talk, 41(5-6):739–762.
Abulimiti Maimaitituoheti, Yong Yang, and Xi-
aochao Fan. 2022. ABLIMET @LT-EDI-
ACL2022: A Roberta based Approach for Homopho-
bia/Transphobia Detection in Social Media. In Pro-
ceedings of the Second Workshop on Language Tech-
nology for Equality, Diversity and Inclusion, pages
155–160, Dublin, Ireland. Association for Computa-
tional Linguistics.
Teresa Marques. 2023. The Expression of Hate in Hate
Speech.Journal of Applied Philosophy, 40(5):769–
787.
Pippa Norris. 2001. Digital Divide: Civic Engagement,
Information Poverty, and the Internet Worldwide.
Communication, Society and Politics. Cambridge
University Press, Cambridge.
Debora Nozza and Dirk Hovy. 2023. The State of Pro-
fanity Obfuscation in Natural Language Processing
Scientific Publications. In Findings of the Associa-
tion for Computational Linguistics: ACL 2023, pages
3897–3909.
Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang,
Yangqiu Song, and Dit-Yan Yeung. 2019. Multilin-
gual and Multi-Aspect Hate Speech Analysis. In
Proceedings of the 2019 Conference on Empirical
Methods in Natural Language Processing and the
9th International Joint Conference on Natural Lan-
guage Processing (EMNLP-IJCNLP), pages 4675–
4684, Hong Kong, China. Association for Computa-
tional Linguistics.
Amado M. Padilla. 1994. Ethnic Minority Scholars,
Research, and Mentoring: Current and Future Issues.
Educational Researcher, 23(4):24–27.
Sara Parker and Derek Ruths. 2023. Is hate speech
detection the solution the world wants? Pro-
ceedings of the National Academy of Sciences,
120(10):e2209384120.
Telmo Pires, Eva Schlinger, and Dan Garrette. 2019.
How Multilingual is Multilingual BERT? In Pro-
ceedings of the 57th Annual Meeting of the Asso-
ciation for Computational Linguistics, pages 4996–
5001, Florence, Italy. Association for Computational
Linguistics.
Mihaela Popa-Wyatt. 2020. Reclamation: Taking Back
Control of Words.Grazer Philosophische Studien,
97(1):159–176.
Clarissa Jane Rajee. 2024. Analyzing Social Values of
Indian English in YouTube Video Comments: A Citi-
zen Sociolinguistic Perspective.Strength for Today
and Bright Hope for Tomorrow Volume 24: 3 March
2024 ISSN 1930-2940, page 9.
Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi,
and Noah A. Smith. 2019. The Risk of Racial Bias
in Hate Speech Detection. In Proceedings of the
57th Annual Meeting of the Association for Computa-
tional Linguistics, pages 1668–1678, Florence, Italy.
Association for Computational Linguistics.
Ella Steen, Kathryn Yurechko, and Daniel Klug. 2023.
You Can (Not) Say What You Want: Using Algos-
peak to Contest and Evade Algorithmic Content Mod-
eration on TikTok.Social Media + Society, 9(3).
Oana Stefania and Diana-Maria Buf. 2021. Hate Speech
in Social Media and Its Effects on the LGBT Com-
munity: A Review of the Current Research. Roma-
nian Journal of Communication & Public Relations,
23(1):47–55.
Ana M. Sánchez-Sánchez, David Ruiz-Muñoz, and
Francisca J. Sánchez-Sánchez. 2024. Mapping Ho-
mophobia and Transphobia on Social Media.Sexual-
ity Research and Social Policy, 21(1):210–226.
Yi Chern Tan and L. Elisa Celis. 2019. Assessing Social
and Intersectional Biases in Contextualized Word
Representations. In Advances in Neural Information
Processing Systems, volume 32. Curran Associates,
Inc.
Alice Tontodimamma, Eugenia Nissi, Annalina Sarra,
and Lara Fontanella. 2021. Thirty years of research
into hate speech: topics of interest and their evolution.
Scientometrics, 126(1):157–179.
Emily van der Nagel. 2018. ‘Networks that work too
well’: intervening in algorithmic connections.Media
International Australia, 168(1):81–92.
Bertie Vidgen and Leon Derczynski. 2020. Direc-
tions in abusive language training data, a system-
atic review: Garbage in, garbage out.PLOS ONE,
15(12):e0243300.
Sidney Wong and Matthew Durward. 2024.
cantnlp@LT-EDI-2024: Automatic Detection
of Anti-LGBTQ+ Hate Speech in Under-resourced
Languages. In Proceedings of the Fourth Workshop
on Language Technology for Equality, Diversity,
Inclusion, pages 177–183, St. Julian’s, Malta.
Association for Computational Linguistics.
Sidney Wong, Matthew Durward, Benjamin Adams,
and Jonathan Dunn. 2023. cantnlp@LT-EDI-2023:
Homophobia/Transphobia Detection in Social Me-
dia Comments using Spatio-Temporally Retrained
Language Models. In Proceedings of the Third Work-
shop on Language Technology for Equality, Diversity
and Inclusion, pages 103–108, Varna, Bulgaria. IN-
COMA Ltd., Shoumen, Bulgaria.
Sidney Gig-Jan Wong. 2023a. Monitoring Hate Speech
and Offensive Language on Social Media. In Fourth
Spatial Data Science Symposium, University of Can-
terbury.
