Computer Vision and Image Understanding: Daksh Balyan, Kashvi Malik, Tanish Gupta, Yash Rakesh, Rachna Narula
Uploaded by
Daksh BalyanComputer Vision and Image Understanding: Daksh Balyan, Kashvi Malik, Tanish Gupta, Yash Rakesh, Rachna Narula
Uploaded by
Daksh BalyanContents lists available atScienceDirect
Computer Vision and Image
Understanding
journal homepage:www.elsevier.com/locate/cviu
CHALLENGES AND APPROACHES IN HINDI HATE SPEECH DETECTION ON SOCIAL
MEDIA PLATFORMS
Daksh Balyana*,Kashvi Malikb*,Tanish Guptac ,YashRakeshd, Rachna Narulae
Department of Computer Science and Engineering,
Bharati Vidyapeeth's College of Engineering, New Delhi – 110063, India
a* b* c
[email protected] [email protected] [email protected]
d e
[email protected] [email protected]
Abstract
heresearchpaperfocusesontheidentificationofhatespeechonsocialmedia,especiallyinthesettingofHindi,
T
whichhasspecialdifficultiesbecauseofthelanguage’scomplexity,lowdataavailability,andcodemixing.Inorder
to handle the enormous volume of textual data produced on social media platforms, the study emphasises the
significance of automated text classification utilising machine learning approaches. By utilisingdevelopmentsin
natural language processing (NLP), techniques including ensemble approaches, deep learning, and traditional
machine learning have greatly enhanced the detection of hate speech.
The study highlights the need of identifying hate speech in Hindi in order to advance online safety, protect
marginalised groups, stop harm from occurring offline, andlessennegativeencountersinreallife.Ittalksabout
using CNNs for deeplearning,TF-IDFforfeatureselection,andXLM-RoBERTaformultilingualtextasmachine
learning techniques to identify hate speech in Hindi. Through tackling these obstacles andutilisingcutting-edge
naturallanguageprocessingmethods,thestudyseekstomaketheinternetamoresecureandwelcomingplacefor
Hindi-speaking users.
Keywords: Deep Learning, XML-RoBERTa, TERABHATE, TF-IDF
1.Introduction
ince socialmediaplatformslikeFacebook,Twitter,andWhatsAppgiveusersarapidandsimplewaytointeract,
S
they are widely used for content creation and information exchange. However, these platforms also serve as
distribution channels for harmful and inflammatory content, which degrades the calibre of online discourse in
addition to their benefits.Hatespeech,whichattacksindividualsorthingsbasedonperceivedidentifyingfeatures
like race, religion, nationality,orsexualorientation,isoneparticularlydestructivesortofsuchcontent.Thebroad
use of social media and the secrecy it offers exacerbate hate crimes. The amount oftextdatainthebigdataera
makes manual classification and processing laborious and susceptible to biases in human judgement, such as
competence and fatigue.
achine learning (ML) techniques can be applied to automated text classification,yieldingpreciseandimpartial
M
outcomes. Notably, developments in machine learning (ML)approaches,includingensemble,deeplearning(DL),
andordinaryML,havesignificantlyimprovedthedetectionofhatespeech.Thisimprovementispartlyattributable
to the amazing advancements achieved in natural language processing, or NLP.
2.Why is it Important to detect Hate Speech
ate speech encompasses a broad variety of remarks that degrade, denigrate, or advocate for violence against a
H
specificpersonorgroup.Itisincompatiblewiththecorevaluesofademocraticsociety,whichincluderespectfor
humanrights,diversity,andinclusivity.Hatespeechisharmfulbecauseitcanincreasetensionsbetweenpeople,stir
up hostility, and in severe situations, even stir up violence.
Detecting hate speech in textual data from social media platforms inHindioranylanguageiscrucialforseveral
reasons:
● P romoting Internet Safety: Hate speech has the abilitytostirupanimosityonline,deterringpeoplefrom
freely expressing their opinions. Hate speech may be identified and reduced to provide a safer online
environment for all users.
● Safeguarding Vulnerable Communities:Hatespeechusuallytargetsmarginalisedgroups,includingracial,
ethnic,andreligiousones.Thesemarginalisedcommunitiescanbeprotectedfromdiscriminationandinjury
via anti-hate speech detection.
● PreventingDamageintheRealWorld:Hatespeechcantransferfrominternetdiscussionboardstophysical
actsofviolenceandbigotry.Earlyidentificationandinterventionwithhatespeechcanhelppreventthese
kinds of situations.
● • Maintaining User Trust: People depend on social media platforms to provide a courteous and secure
environment. These platforms can maintain user confidence and draw in new users by recognising and
suppressing hate speech.
● Reducing Toxicity:Thepresenceofhatespeechmakesonlinediscussionsmoretoxic.Byidentifyingand
correcting hate speech, forums can diffuse negative energy and promote more constructive discourse.
Figure 1:
s depicted in Figure 1. Genocideservesasafertilegroundforhatespeechbecauseitdehumanisesthegroupsit
A
targets, propagates propaganda demonising them, and uses them asscapegoatsforsocialissues,allofwhichsow
division and hatred in society.
3.Categories of Hate Speech in Hindi Language
N
S CATEGORIES EXAMPLE OF HATE TARGET
O
1) Race काले,नीच,दfलत,अछू त
2) Behavior ढkला,छु ई मुई
3) Physical बौना,ग डा,ल
कड़ब घा,fचकना
4) Sexual Orientation fहजरा
5) Class गरkब,अमीरजादा
6) Gender मद,जनानी
7) Ethnicity fचकं ,मǐलू
8) Disability लंगड़ा,अंधा,बैहरा,भग हा
9) Religion धम fवरोधी
10) Others शराबी,उथला यिRत
I t is imperative to remember that hate speech that disparagesindividualsorgroupsonthebasisoftheselabelsis
harmfulandcanhavedireconsequences,includingviolenceandsocialexclusion.Everyoneneedstobetreatedwith
dignity and respect, regardless of where they are from or what group they belong to.
This explanation emphasisestheimportanceofpromotingdiversityandrespectwhileacknowledgingthenegative
impacts of hate speech, remaining unbiased, and avoiding using judgmental language.
atespeechpromotesaggressivebehaviourbynormalisinghostilitytowardsparticulargroups,dehumanisingthem
H
to justify violent behaviour, and endorsing the use of force to resolve disputes. It escalates tensions, radicalises
individuals,andcreatesanunstableatmospherethatraisesthepossibilityofviolence.Hatespeechhastheabilityto
inflictmiseryintherealworldbyexposingindividualstothedetrimentalimpactsofviolenceonaregularbasisand
by using influential figures.
Hate speechfeedsprejudicebyfosteringnegativeviewsandassumptionsaboutspecificgroups,whichcanleadto
unequal treatment and exclusion. By dehumanising and demonising the people it targets, hate speech helps to
legitimise discriminatory actions and spread prejudice. It can affect policies and practices that disadvantage or
marginalise specific communities and cancontributetothedevelopmentofaclimatewhereprejudiceistolerated.
Additionally, hate speech can widen rifts in society, undermine efforts to promote equality and inclusivity, and
disrupt social cohesion. All things considered, hate speech promotes discrimination by upholding unfavourable
beliefs and attitudes that back up systemic injustice and inequality.
atespeechhasthepowertoprovokeactsofprejudicebypromotingnegativestereotypesandbiasesagainstspecific
H
groups,whichcaninfluencehowpeopleseeandinteractwiththemembersofsuchgroups.Hatespeechcanincite
discrimination in a variety of ways, such aswhenittargetscertainindividualsduetotheirperceivedidentityand
leads to harassment, violence, and destruction. It may also manifest as prejudice in employment, housing, or
educational opportunities. In addition to causingimmediatemisery,thesebiassedbehavioursexacerbatestructural
injusticesanddeepensocietaldivisions.Hatespeechmustbehandledinordertoputanendtoactsofdiscrimination
and create a society that is more diverse and equitable.
4.Difficulties with Hindi Hate Speech Detection
hilesignificantstrideshavebeenmadeinhatespeechdetectioninlanguageslikeEnglish,thesamecannotbesaid
W
for Hindi, one of the most widely spoken languages globally. This research endeavour confronts several unique
challenges:
❖ L imitedAccesstoData:TherearecomparativelyfewresourcesavailableforHindilanguageprocessingas
comparedtolanguageslikeEnglish.Thelackofannotateddatasetsmakesitdifficulttotrainandimprove
machine learning frameworks for recognising hate speech in Hindi.
❖ The Complexity of Language: Hindi hasasophisticatedscriptandarichmorphology.Compoundterms,
colloquial idioms, and regional differences add another level of complication to the challenge of
recognising hate speech.
❖ Situational Ambiguity: Like many other languages, Hindifrequentlydependsoncontextualsignalstobe
understood correctly. Understanding cultural, social, and historicalbackgroundsindetailisnecessaryfor
the detection of hate speech, and this may not always be easy for automated systems to do.
❖ MergingCodes:SocialmediaconversationsinHindimayinvolvecode-mixing,whichistheutilisationof
many languages—including English—in a single sentence or remark. Because of this, itisconsiderably
harder to detect hate speech because models must be adept at managing multilingual information.
❖ Inlightofthesechallenges,thisresearchendeavourstoclosethegapintheidentificationofhatespeechby
concentratingontheHindilanguage.Byaddressingtheseuniquehurdles,weaimtocontributetoamore
secure and welcoming online community for people who speak Hindi.
5.Challenges faced during the implementation in Hindi languages
hen it comes to identifying hate speech in Hindi, there are many unique challenges because of the complexity of
W
the language, the lack of readily available data, and the prevalence of code mixing. The following is a breakdown
of these challenges:
1. Language Complexity: Hindi has a rich morphology, a huge vocabulary, and a sophisticated syntax. It
can be challenging to accurately comprehend and classify content due to its complexity, especially when
dealing with the informal or colloquial language that is commonly used on social media.
2. Low Data Availability: Compared to languages like English, there is considerably less labeled data
available for Hindi hate speech identification. It is difficult to train machine learning models with
little annotated data since these models require large amounts of data to learn effectively.
3. Code mixing, often known as blending two or more languages into a single statement or debate, is a
common practice in Hindi writing, especially in informal online communication. Standard natural language
processing (NLP) models may have trouble reading or interpreting code-mixed text because they are
trained on single-language datasets. These models could become confused by this phenomenon.
4. Contextual Understanding: One must be aware of the context of a statement in order to properly categorize
it as hate speech. The way a text is interpreted in Hindi can be significantly impacted by contextual details,
c ultural allusions, and regional differences. As a result, developing trustworthy hate speech detection
models that account for these variances is challenging.
6 .Ethical Considerations In Automated Hate Speech Detection
In automated hate speech identification, ethical considerations are crucial to preserve cultural sensitivity and
minimize potential biases, especially when working with languages such as Hindi. The following are important
points to keep in mind:
1. C ultural Sensitivity: Automatic systems for detecting hate speech must be aware of subtle linguistic
variations as well as cultural diversity. This means that in order to avoid mistaking kind words for hate
speech, one must be cognizant of linguistic diversity, regional variations, and cultural context.
2. Bias Mitigation: Automated systems may inherit biases from the training data. The identification of hate
speech in Hindi may be subject to biases because of the tiny and perhaps uneven dataset. Techniques like
data augmentation, the collection of diverse datasets, and bias identification and mitigation algorithms
are essential to reducing these biases.
3. Explainability and Transparency: Automated systems need to be transparent about their decision-making
processes and able to provide an explanation for their hate speech identification decisions. This helps users
understand why a certain piece of content was reported and allows for the identification of any biases or
errors in the system.
4. Human Testing: Hiring human reviewers can help minimize biases and ensure that automated systems are
making accurate conclusions when cultural context plays a significant role in the process.
5. Privacy and Data Protection: Hate speech detection systems must abide by strict privacy and data
protection rules, especially when handling sensitive user information. Ensuring data encryption,
anonymization, and user consent are necessary to maintain user confidence.
7.Literature Review
Ref. METHODOLOGY LIMITATION FINDINGS
[1] sed lexicon-based approach and
U imited to Hindi, may not apply to T
L ABHATE is the first
XLM-RoBERTa for tweet other languages, cultural/regional target-based Hindi hate speech
classification. variations not considered. detection dataset.
Over 10,000 tweets have been
annotated in this collection; each
one has been marked with the
group that the hate speech is
directed against and is
categorized as either hate speech
or not.
The accuracy of Hindi hate
speech detection methods is
improved by the TABHATE
dataset.
[2] re-processing: Case folding,
P volving attitudes and context.
E ate speech recognition systems
H
tokenization, and punctuation Closed loop systems aiding have several challenges,
removal.Feature identification: evasion. including the need to handle
Unigram to 5-gram TF-IDF Ethical concerns with user data. multiple languages and cultures,
counts.Multi-view SVM model: Future hate speech identification the challenge of classifying hate
Utilizing different feature types. complexity. speech, and the absence of
annotated data.
The paper offers a machine
learning technique that combines
multiple characteristics to
identify hate speech in order to
address these problems.
[3] ata collection: Public hate speech
D imited dataset: Relied on a
L he report claims that machine
T
tweets gathered. Text specific hate speech dataset. learning is the foundation for the
pre-processing: Irrelevant Language dependency: Focused on method for automatically
information removed. Feature English hate speech detection. identifying hate speech.
engineering: Text converted to Generalizability: Findings may not The system uses a range of
numerical vectors using software. apply to all platforms. variables, including text, audio,
Data splitting: Dataset divided into Performance evaluation: Metrics and visual aspects, to identify
training and test sets. and classifiers were specific. hate speech.
A high accuracy of the system is
demonstrated by an evaluation
with a Hindi hate speech dataset.
[4] ata Collection: Obtained data
D imited Data Availability:
L he publication presents the new
T
from reliable sources. Restricted by data availability. dataset HateCheckHIn, which is
Preprocessing: Cleaned and Time Constraints: Limited time for used to evaluate models for
formatted data. Feature Extraction: data work. Scope: Focused on detecting hate speech in Hindi.
Extracted key features. Model specific variables. The collection contains over
Selection: Choose suitable models 10,000 labeled tweets, with labels
for analysis. designating whether or not the
content constitutes hate speech.
The research also presents a
novel hate speech recognition
algorithm that shows
state-of-the-art performance
when tested on the
HateCheckHIn dataset.
[5] ata collection: Twitter tweets
D nbalanced dataset: Challenges in
U he research proposes a novel
T
gathered with specific hashtags. equal representation of hate and Arabic BERT-Mini model,
Pre-processing: Removed non-hate tweets. dubbed ABMM, for the
punctuation, normalized Arabic Lack of additional features: identification of hate speech on
text, eliminated usernames, URLs, Ignored factors like emoji social media.
and hashtags. descriptions. The model achieves
Model development: Created Interpretability:ABMMlacks state-of-the-art performance
Arabic BERT-Mini Model explanationsforclassification when evaluated on three Arabic
(ABMM) for hate speech detection. decisions. hate speech datasets.
For real-time systems that
identify hate speech, ABMM is a
valuation: Compared ABMM
E ardware constraints: Limited by
H p romising option due to its
withtraditionalMLmodelsand memory and CPU for deep portability and efficiency.
state-of-the-art approaches. learning layers.
[6] ata Collection: Gathered relevant
D imited Data Availability:
L he goal of arHateDetector, a
T
data. Constrained by data availability. brand-new collection of standard
Preprocessing: Cleaned and Time Constraints: Research and dialectal Arabic tweets, is to
standardized data. conducted within a specific identify hate speech.
Model Selection: Choose timeframe. A novel deep learning model is
appropriate models. Resource Limitations: Limited constructed that yields
Model Training: Trained models computational resources. cutting-edge results on the
using preprocessed data. arHateDetector dataset.
The model can be applied to
real-time hate speech detection
applications because to its
portability and efficiency.
[7] ix transformer models for hate
S imited computational resources
L ate speech is a serious problem
H
speech detection. for model size. in Arabic-speaking countries, so
Two ensemble methods to combine Dataset focus on COVID-19 it's imperative to have effective
models. disinformation restricts tools for spotting it.
Cross-validation for performance generalizability. Transformer-based models have
evaluation. Models trained on Arabic text been shown to perform
Best model selection based on limit language applicability. exceptionally well in tasks
F1-score. Evaluation metric focuses solely related to natural language
on F1-score for the positive class. processing, including the
detection of hate speech.
Transformer-based models can
function more effectively when
the predictions of multiple
models are combined using
ensemble methods.
[8] ata Collection: Gathered tweets
D ingleexpertannotatorpertweet.
S nnotated AraCOVID19-MFH
A
using specific keywords. Smalldataset(10,828tweets)may dataset including 10,828 Arabic
Data Annotation: Expert-labeled impact model performance. tweets for detecting fake news
tweets. FocusedonArabictweets,limiting and hate speech.
Preprocessing: Basic text generalizability. The dataset was used to train and
preprocessing applied. evaluate models, with positive
Model Training: Pretrained results in classification tasks.
transformer models used for Pretraining models with
classification. COVID-19 data improved their
performance, particularly in hate
speech and fake news
identification.
[9] reprocess text data: Remove
P imited Hindi hate speech
L he paper develops a model for
T
URLs, usernames, punctuation. datasets. detecting hate speech in
Translate emoticons to Hindi Imbalanced distribution of hate low-resource Hindi using BERT
descriptions. and non-hate classes. and a Deep Convolution Neural
se pretrained BERT encoder for
U ddressed imbalance with
A etwork. In comparison to the
N
contextualized embeddings. oversampling. baseline BERT model, the model
Implement CNN with parallel Evaluation based on f1-score due improved its f1-score
convolution filters. to dataset imbalance. significantly. Future research
may explore combining other
BERT variations with different
deep learning classification
algorithms to improve hate
speech detection.
[10] thorough review of hate speech
A ate speech recognition systems
H he paper discusses current
T
detection literature was undertaken may fail to recognise contextual advances in hate speech
using rigorous criteria to select nuances, resulting in incorrect identification, including the use
relevant publications. identification of hate speech. The of text and image fusion
The study carefully examined study addresses biases in data and techniques. Various models that
selected articles to extract algorithms, which affect fair use traditional machine learning
important components such as detection results. It also and deep learning methodologies
techniques, datasets, performance emphasises the difficulty of have considerably improved hate
measures, and noteworthy findings multilingualism, where language speech recognition. Recent
in the field of hate speech variations impede reliable hate research has focused on
identification. speech identification across multi-task learning, Bayesian
diverse linguistic contexts. techniques
[11] he paper's methodology includes
T he paper's limitations include a
T he paper's findings emphasise
T
a literature assessment on hate probable lack of coverage of the marginalisation of minority
speech detection techniques, data developing hate speech detection groups through hate speech,
collection and normalisation from algorithms, a concentration on present a taxonomy of automatic
social media sites, and automatic textual data rather than multimodal hate speech identification, and
hate speech identification using inputs, and the need for further describe metaheuristic algorithms
classification techniques such as investigation of issues in detecting and works in multilingual and
machine learning and deep hate speech in multilingual multimodal hate speech
learning. situations. detection.
8.Comparative Analysis Of Various Techniques Used
.1.XLM RoBERTa -
8
When it comes to identifying hate speech, XLM-RoBERTa canbeaveryusefulinstrument.XLM-RoBERTaisa
transformer-based model that has undergone substantial pretraining on a large volume of textualdatainmultiple
languages.
Jobs containing multilingual text benefit from this model’s specific training to handle languages with various
qualities.Itsfine-tuningflexibilityallowsadaptiontodownstreamapplicationswithalesserrequirementforlabelled
data, compared to training from scratch. Because XLM-RoBERTa can generate accurate and contextually rich
representations in several languages, it is a valuable tool for various natural language processing applications.
UsingXLM-RoBERTaforHateSpeechDetectionincludescreatingadatasetthatincludesexamplesofhatespeech
aswellasregulartextandnon-hatespeechexamples.Forpreprocessing,textdataistokenizedsothatwordsegments
orsubwordscanbeinsertedintoXLM-RoBERTa.Thenthistextdataisconvertedintoastructurethatthemodelis
abletocomprehend.TheXLM-RoBERTamodelisinitialisedusingpretrainedweights.Byincludingaclassification
layer to identify whether or not a given input contains hate speech, the XLMRoBERTa model can be improved. The
e ntiremodelofhatespeechisfine-tunedusingthetaggeddataset.Themodel’sperformanceisevaluatedusingthe
test set. Common evaluation metrics for binary classification tasks like hate speech detection include accuracy,
precision, recall, F1-score, and area under the ROC curve (AUC).
.2.TD-IDF-
8
Term frequency-inverse document frequency, or
TF-IDF , is a method that evaluates a word’s importance in a document by looking at how often it occurs in a
collection of documents. It is often used in text representation machine learning models.
The TF-IDF approach offers several advantages in text analysis. To begin with, it performs an excellent job of
feature selection by identifying and selecting key terms that enhancethedocument’scontentandenableadeeper
comprehension oftheunderlyingfacts.Additionally,TF-IDFhelpsreducedimensionalitybyfocusingonthemost
relevant features. This is particularly useful when handling high-dimensional data, which can cause computing
issues.Becauseofthis,itmaybeusedtoevaluatelargedatasetswithalargenumberofdocuments,whichenhances
scalability and practicality in real-world applications.
8.3.CNN or Convolutional Neural Network-
he effectiveness of a CNN data analysis model is critical for many applications. Efficient CNNs have lower
T
memory footprints, faster inference times, and shorter training periods—all of which are critical for real-time
applications and resourceconstrained scenarios. Techniques like hardware optimisation, quantization, and model
pruning reduce memory usage and boost computing performance. Additionally,scalability,energyefficiency,and
resilience are important considerations. CNN model optimisation involves striking a balance between accuracy,
speed, and resource consumption based on specific deployment scenarios and performance requirements.
Convolutional Neural Networks (CNNs) have several advantages, especially when processing text and image
data—two typesofdatathatarecommonlyusedinapplicationsforthedetectionofhatespeech.FeatureLearning
andSpatialHierarchiesaretheprincipaladvantagesofCNN.CNNsareabletoautomaticallylearnrelevantqualities
fromrawdata,suchasedges,shapes,andtexturesinimagesorn-gramsandpatternsintext,withouttherequirement
forhumanfeatureextraction,whichiscalledFeatureLearning.Bylearninghierarchicaldatarepresentations,CNNs
are abletorecognisespatialhierarchiesoffeatures.Whilelowerlayersinanimagemaylearnbasicattributeslike
borders, higher layers may learn morecomplexfeatureslikefullobjectsorportionsofobjects.Lowerlayersmay
pick up word-level characteristics from text, while higher layers might pick up sentence- or document-level
information.
.4.Multiview SVM Model -
8
This is a classification algorithm. The efficiency and effectiveness of a Multi-view SVM model for hate speech
detection depend on various factors, including the quality of the data, the choice of views, feature extraction
techniques, model architecture, and the complexity of the hate speech detection task itself.
SVMsperformexceptionallywellwhentheappropriatekernelfunctions—suchaspolynomial,radialbasisfunction
(RBF), or sigmoid kernels—areappliedtodatathatisbothlinearlyandnon-linearlyseparable.SVMsareflexible
enoughtoanticipatecomplexrelationshipsbetweenfeatures,makingthemsuitableforawiderangeofclassification
and regression tasks. SVMs are resistant to overfitting and can better generalise to fresh data since they may
maximisethemarginbetweenclasses.SVMsalsoperformwellinhigh-dimensionalspacesandarenothamperedby
the curse of dimensionality, which makes them suitable for datasets with a large number of characteristics.
9..Hate Speech Data On Social Media In Hindi Language
Figure 2: USERS ACTIVE ON SOCIAL MEDIA
his pie chart shows the distribution of active users across the four main social media platforms in the context of
T
detecting hate speech. With 40% of users, Facebook is the most popular platform, suggesting that its enormous user
base makes it a potentially valuable source of data for the identification of hate speech. Instagram, which has 33% of
its users, represents a somewhat smaller but significant share. It is a visual content-rich network that may create
particular difficulties when it comes to identifying hate speech. Twitter offers a real-time text-based environment
with a 10% stake, which may provide useful information regarding hate speech trends and patterns. A major section
of the online community is represented by YouTube, which has 17% of users and offers video footage that could add
more context for hate speech detection. The distribution of active users across different platforms, taken as a whole,
highlights the significance of taking into account multiple platforms in hate speech detection in order to capture a
varied variety of material and user behaviors, guaranteeing a more thorough and successful approach to tackling this
crucial issue.
Figure 3: HATE CONTENT(%)
heaccompanyingpiechartshowsthedistributionofhatecontentacrossfourkeyplatformsinthecontextofsocial
T
media hate speech detection. With 37% of the total, Twitter is the platform with the highest percentage of hate
content.ThisdiscoveryhighlightsTwitter’spotentialasamajorconduitforhatespeech,presumablybecauseofits
real-time nature and the ease with which content can be disseminated and magnified. With25%ofhatecontent,
Instagram comes in close second, showing that hate speech can also spread on visual media. Facebook has the
greatestuserpopulation,yetits20%hatecontentrateislowerthanthatofotherplatforms,indicatingthatitscontent
filtering practices maybemoresuccessful.Duetoitsstrongerproceduresoncontentmonitoringandthenatureof
video content, which may make itmoredifficultforhatespeechtospread,YouTubehasthelowestpercentageof
hate content (18%). These results emphasize the significance of customized methodsforidentifyingandfiltering
hate speech on various social media platforms, taking into account their particular characteristics and difficulties.
10.Objectives
➢To design a meticulously curated and annotated datasetwith a focus on target-based hate speech detection,
particularly within the context of the Hindi language.
➢Develop a Robust Hate Speech Detection Algorithm forHindi Text
➢Implement Advanced NLP Techniques for Hindi
➢Apply advanced NLP techniques, including sentimentanalysis and contextual understanding, specifically
tailored to the complexities of the Hindi language.
➢Evaluate Model Performance and Fine-tune as Necessary
➢Rigorously assess the hate speech detection model’s accuracy, sensitivity, and specificity using a variety of
evaluation metrics. Make necessary adjustments to improve its effectiveness.
11.Workflow Diagram
Figure 4: Workflow Diagram
1 2.Future Directions
Future work on automated hate speech detection will prioritise multilingualism. To identify hate speech across
languages, models must be developed. Improving contextual understanding of hate speech, such as detecting
sarcasm and irony, is a top priority. Providing comprehensive explanations of the detection system's decisions is
critical for building user confidence and driving future progress. To detect hate speech, it's important to design
resilient models that can withstand adversarial attacks. Real-time detection of hate speech is crucial for improving
internet safety and avoiding harm. Developing multimodal algorithms that can recognise hate speech in text,
graphics, and videos is vital. Future study should address bias in hate speech detection systems to achieve accurate
and equitable outcomes. Advancements in technology offer opportunities for automated hate speech identification
in the future. One potential use is a real-time social media monitoring programme that highlights hate speech for
human moderators to review. Another potential application is a chatbot that provides tools to help consumers
understand the harm caused by hate speech during online conversations. Automatic hate speech detection in
messaging apps and online forums can help limit the spread of hate speech and toxic environments. As technology
progresses, new applications for automatically detecting hate speech will emerge to foster safer and more inclusive
o nline communities. Speech can be labelled to identify online sexual violence, suicidal thoughts, and other topics
beyond hate speech alone.
Some research studies (Khatua et al.2018) have focusedon identification of gender-based violence on Twitter
using #Metoo movement, which were done using Deep Learning based lexical approaches.
Some sections of the society like the lower caste, dalits, LGBTQ+ community has always been discriminated
against. A study conducted in (Khatua et al.2019)proposed ascpect extraction method to comprehend the root cause
of such discrimination against minorities. Also, textual content analysis techniques (Ji et al.2020) like
lexicon-based filtering and wordcloud visualisation and feature engineering techniques like tabular, textual, and
affective features for detecting suicidal ideation on social media platforms. Another relevant problem on social
media platforms is Fake News detection (Kansara & Adhvaryu,2022). This can be solved using cutting-edge
learning strategies, interpretable intention understanding, temporal detection and and proactive conversational
intervention. Another study (Jafaar and Lachiri,2023)combined all four multimodal methods, which included
audio, video and text. This analysed how acoustic, visual and textual features are combined and how they have an
impact on the fusion process and level of aggression.
13.Conclusion
o build user trust, safeguard vulnerable communities, stop harm in the real world, promote internet safety, and
T
reduce the toxicity of online interactions, it is essential to identify hate speech in languages like Hindi. However, a
number of challenges exist when it comes to identifying hate speech in Hindi, including a dearth of easily
accessible data, language complexity, situational ambiguity, and code mixing.
Despite these obstacles, recent advances in machine learning, particularly in the field of natural language processing,
provide hopeful avenues for progress. Textual data can be used to determine hate speech with the aid of algorithms
such as CNNs, TF-IDF, and XLM-RoBERTa, which leverage deep learning architectures and advanced feature
extraction techniques.
By addressing these problems and employing state-of-the-art machine learning techniques, our goal is to contribute
to making the internet a more secure and inclusive space for Hindispeaking populations.
References
[ 1] D. Sharma, V. K. Singh, and V. Gupta, "TABHATE: A Target-based Hate Speech Detection Dataset in
Hindi," Research Article, Banaras Hindu University and Jindal Global Business School, O.P. Jindal Global
University, April 20, 2023. DOI:https://europepmc.org/article/ppr/ppr648319.
[ 2] Dr. A.B. Pawar, Dr. M. A. Jawale, Pranav Gawali, and P. William, "Challenges for Hate Speech
Recognition System: Approach based on Solution," in 2022 International Conference on Sustainable
Computing and Data Communication Systems (ICSCDS), doi:
https://ieeexplore.ieee.org/abstract/document/9760739.
[ 3] *P. William, Dr. A. B. Pawar, Ritik Gade, and Dr. M. A. Jawale, "Machine Learning based Automatic
Hate Speech Recognition System," in Proceedings of the International Conference on Sustainable Computing
and Data Communication Systems(ICSCDS-2022),
doi:-https://www.researchgate.net/publication/363049790_Machine_Learning_based_Automatic_Hate_Speech_Rec
ognition_System.
[ 4] M. Das, P. Saha, B. Mathew, and A. Mukherjee, "HateCheckHIn: Evaluating Hindi Hate Speech
Detection Models," in Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022),
Marseille, 20-25 June 2022,https://arxiv.org/abs/2205.00328.
[ 5] M. Almaliki, A. M. Almars, I. Gad, and E.-S. Atlam, "ABMM: Arabic BERT-Mini Model for
Hate-Speech Detection on Social Media," Electronics, vol. 12, p. 1048, Feb. 2023, doi:
https://www.mdpi.com/2079-9292/12/4/1048.
[ 6] R. Khezzar, A. Moursi, and Z. A. Aghbari, "arHateDetector: detection of hate speech from standard and
dialectal Arabic Tweets," vol. X, no. X, pp. X, 2023, doi:
https://link.springer.com/article/10.1007/s43926-023-00030-9.
[ 7] A. F.M.d.Paula,I.Bensalem,P.Rosso,andW.Zaghouani,"TransformersandEnsemblemethods:A
solution for Hate SpeechDetectioninArabiclanguages,"inCEURWorkshopProceedings,2022,vol.1,ISSN
1613-0073, DOI:https://arxiv.org/abs/2303.09823.
[ 8] M.S.H.AmeurandH.Aliane,"AraCOVID19-MFH:ArabicCOVID-19Multi-labelFakeNews&Hate
Speech Detection Dataset," Procedia Computer Science, 2021. Available: www.sciencedirect.com. [Online].
Available: doi:https://www.sciencedirect.com/science/article/pii/S1877050921012059.
[ 9] S. Shukla, S. Nagpal, and S. Sabharwal, "Hate Speech Detection in Hindi language using BERT and
Convolution Neural Network," Netaji Subhas University of Technology, Delhi, India.
doi:-https://ieeexplore.ieee.org/document/10037649.
[ 10] Gandhi, A., Ahir, P., Adhvaryu, K., Shah, P., Lohiya, R., Cambria, E., Poria, S., Hussain, A., "Hate
speech detection: A comprehensive review of recent works," Expert Systems, vol. 41, no. 4, p. e13562, 2024,
doi:https://doi.org/10.1007/s00530-023-01051-8
[ 11] A. Chhabra and D.K. Vishwakarma, "A literature survey on multimodal and multilingual automatic hate
speech identification," Multimedia Systems, vol. 29, no. 4, pp. 1203–1230, 2023, doi:
https://link.springer.com/article/10.1007/s00530-023-01051-8
[ 12]Ma, Z., Yao, S., Wu, L., Gao, S., & Zhang, Y. (2022). Hateful memes detection based on multi-task learning.
Mathematics, 10(23), 4525
[ 13] Miok, K., Skrlj, B., Zaharie, D., & Robnik- ˇ Sikonja, M. (2022). To ban or not to ban: Bayesian
attention networks for reliable hate speech detection. ˇ Cognitive Computation, 14, 353–371.
[ 14] Montariol, S., Riabi, A., & Seddah, D. (2022). Multilingual auxiliary tasks training: Bridging the gap
between languages for zero-shot transfer of hate speech detection models. arXiv preprint arXiv:2210.13029.
[ 15] Mozafari, M., Farahbakhsh, R., & Crespi, N. (2022). Cross-lingual few-shot hate speech and offensive
language detection using meta learning. IEEE Access, 10, 14880–14896.
[ 16] Mridha, M. F., Wadud, M. A. H., Hamid, M. A., Monowar, M. M., Abdullah-Al-Wadud, M., & Alamri,
A. (2021). L-Boost: Identifying offensive texts from social media post in Bengali. IEEE Access, 9,
164681–164699.
You might also like
- Hate Speech Detection - Challenges and Solutions - PLOS ONENo ratings yetHate Speech Detection - Challenges and Solutions - PLOS ONE9 pages
- BDA Minor Specialization Literature Review IEEE FormatNo ratings yetBDA Minor Specialization Literature Review IEEE Format5 pages
- Detecting and Visualizing Hate Speech in Social Media: A Cyber Watchdog For Surveillance-Modha2020No ratings yetDetecting and Visualizing Hate Speech in Social Media: A Cyber Watchdog For Surveillance-Modha202011 pages
- A Review of Challenges in Machine Learning Based Automated Hate Speech DetectionNo ratings yetA Review of Challenges in Machine Learning Based Automated Hate Speech Detection9 pages
- Hate Speech Detection: Challenges and Solutions: A1111111111 A1111111111 A1111111111 A1111111111 A1111111111No ratings yetHate Speech Detection: Challenges and Solutions: A1111111111 A1111111111 A1111111111 A1111111111 A111111111116 pages
- A Survey On Automatic Detection of Hate Speech in TextNo ratings yetA Survey On Automatic Detection of Hate Speech in Text30 pages
- Detection of Hate Based Political SpeechNo ratings yetDetection of Hate Based Political Speech5 pages
- Combating Online Hate A Comparative Study On Identification of Hate Speech and Offensive Content in Social Media TextNo ratings yetCombating Online Hate A Comparative Study On Identification of Hate Speech and Offensive Content in Social Media Text6 pages
- Ensuring Safety in Digital Spaces Detecting Code Mixe 2025 Data KnowledgeNo ratings yetEnsuring Safety in Digital Spaces Detecting Code Mixe 2025 Data Knowledge15 pages
- Enhancing Hate Speech Detection in The Digital Age A Novel Model Fusion Approach Leveraging A Comprehensive DatasetNo ratings yetEnhancing Hate Speech Detection in The Digital Age A Novel Model Fusion Approach Leveraging A Comprehensive Dataset12 pages
- Kec Ai Gryffindor Dravidianlangtech Naacl 2025No ratings yetKec Ai Gryffindor Dravidianlangtech Naacl 20257 pages
- Multi-Class Hate Speech Detection in The Norwegian Language Using FAST-RNN and Multilingual Fine-Tuned TransformersNo ratings yetMulti-Class Hate Speech Detection in The Norwegian Language Using FAST-RNN and Multilingual Fine-Tuned Transformers22 pages
- Detection of Hate Speech Using BERT and Hate Speech Word Embedding With Deep ModelNo ratings yetDetection of Hate Speech Using BERT and Hate Speech Word Embedding With Deep Model23 pages
- Misogynistic Attitude Detection in YouTube Comments and R 2025 Computer SpeeNo ratings yetMisogynistic Attitude Detection in YouTube Comments and R 2025 Computer Spee16 pages
- Overview of The HASOC Subtrack at FIRE 2023: Identification of Tokens Contributing To Explicit Hate in English by Span DetectionNo ratings yetOverview of The HASOC Subtrack at FIRE 2023: Identification of Tokens Contributing To Explicit Hate in English by Span Detection8 pages
- Hate Speech Detection in Twitter Using Natural Language ProcessingNo ratings yetHate Speech Detection in Twitter Using Natural Language Processing7 pages
- Hatexplain: A Benchmark Dataset For Explainable Hate Speech DetectionNo ratings yetHatexplain: A Benchmark Dataset For Explainable Hate Speech Detection12 pages
- Detection of Hate Speech Using BERT and Hate Speech Word Embedding With Deep ModelNo ratings yetDetection of Hate Speech Using BERT and Hate Speech Word Embedding With Deep Model23 pages
- A Survey On Automatic Online Hate Speech Detection in Low-Resource LanguagesNo ratings yetA Survey On Automatic Online Hate Speech Detection in Low-Resource Languages34 pages
- Hatemonitors: Language Agnostic Abuse Detection in Social MediaNo ratings yetHatemonitors: Language Agnostic Abuse Detection in Social Media8 pages
- HateXplain: A Benchmark Dataset For Explainable Hate Speech DetectionNo ratings yetHateXplain: A Benchmark Dataset For Explainable Hate Speech Detection9 pages
- A Comprehensive Review On Automatic Hate Speech Detection in The Age of The TransformerNo ratings yetA Comprehensive Review On Automatic Hate Speech Detection in The Age of The Transformer25 pages
- On Early Detection of Hallucinations in Factual Question AnsweringNo ratings yetOn Early Detection of Hallucinations in Factual Question Answering11 pages
- Paper 57-Offensive Language Detection On Social MediaNo ratings yetPaper 57-Offensive Language Detection On Social Media8 pages
- A Multilingual Evaluation For Online Hate Speech DetectionNo ratings yetA Multilingual Evaluation For Online Hate Speech Detection22 pages
- Buy Cosmetics & Beauty Products Online From Nykaa, The Online Shopping Beauty Store. Browse Makeup, Health Products & MoNo ratings yetBuy Cosmetics & Beauty Products Online From Nykaa, The Online Shopping Beauty Store. Browse Makeup, Health Products & Mo1 page
- Diabetic Retinopathy Detection Using Deep Learning Techniques - A Comprehensive StudyNo ratings yetDiabetic Retinopathy Detection Using Deep Learning Techniques - A Comprehensive Study13 pages
- Diabetic Retinopathy Detection Using Deep Learning Techniques - A Comprehensive StudyNo ratings yetDiabetic Retinopathy Detection Using Deep Learning Techniques - A Comprehensive Study8 pages
- Second Language Acquisition (Contrastive Analysis Hypothesis)No ratings yetSecond Language Acquisition (Contrastive Analysis Hypothesis)9 pages
- Communication and Motivation in Sport Management Physical EducationNo ratings yetCommunication and Motivation in Sport Management Physical Education4 pages
- An Analysis of Lexical Cohesion Found in Red SongNo ratings yetAn Analysis of Lexical Cohesion Found in Red Song9 pages
- Attitudes Towards FB Advertising Questionnaire0% (1)Attitudes Towards FB Advertising Questionnaire3 pages
- 2024 English Month Celebration Project ProposalNo ratings yet2024 English Month Celebration Project Proposal4 pages
- How To Write in Gallifreyan - Loren ShermanNo ratings yetHow To Write in Gallifreyan - Loren Sherman33 pages