Hate Speech Research Articles

Leveraging Unsupervised Task Adaptation and Semi‐Supervised Learning With Semantic‐Enriched Representations for Online Sexism Detection

ABSTRACTOver the past decade, the proliferation of hateful and sexist content targeting women on social media has become a concerning issue, adversely affecting women's lives and freedom of expression. Previous efforts to detect online sexism have utilized monolingual ensemble transformers combined with data augmentation techniques that incorporate related‐domain data, such as hate speech. However, these approaches often struggle to capture the full diversity and complexity of sexism due to limitations in the size and quality of training data. In this study, we introduce a novel sexism detection system that employs in‐domain unlabeled data through unsupervised task‐adaptation techniques and semi‐supervised learning, using an efficient single multilingual transformer model. Additionally, we incorporate a Sentence‐BERT layer to enhance our system with semantically meaningful sentence embeddings. Our proposed system outperforms existing state‐of‐the‐art methods across all tasks and datasets, demonstrating its effectiveness in detecting and addressing sexism in social media text. These results underscore the potential of our approach, providing a foundation for further research and practical applications.

A Sentiment Analysis of Hate Speech in Philippine Election-Related Posts Using BERT Combined with Convolutional Neural Networks and Model Variations Incorporating Hashtags and ALL-CAPS

As the number of people who use X continually increases, the same thing is true for hate speech. A pressing need exists for automatic detection of posts that promote hate speech. The datasets gathered and validated from the base study were used to categorize posts as either hate or non-hate and classify them as positive, negative, or neutral using Conventional Neural Networks. The partitioning of the labeled data into training and testing sets adhered to a ratio scheme: 70%-30%, 80%-20%, and 90%-10%. The model of this study, BERT-CNN, had an overall better performance than the base study, fastText CNN. Notably, among the three splits, the BERT-CNN model for binary classification without the features of Hashtags and ALL-CAPS with the 90:10 split achieved the best performance with an accuracy of 93.55%, precision of 93.59%, and F1-score of 93.55%. For multi-label classification, the BERT-CNN model demonstrated its optimal performance when incorporating hashtags, specifically with the 90:10 split, achieving an accuracy of 69.14%, precision of 68.44%, recall of 68.40%, and an F1-score of 67.41%. The innovative use of BERT word embeddings paired with CNN proved to excel in classifying Philippine election-related posts as hate or non-hate.

Political-RAG: using generative AI to extract political information from media content

ABSTRACT In the digital era, media content is crucial for political analysis, providing valuable insights through news articles, social media posts, speeches, and reports. Natural Language Processing (NLP) has transformed Political Information Extraction (IE), automating tasks such as event extraction and sentiment analysis. Traditional NLP methods, while effective, are often task-specific and require specialized expertise. In contrast, Large Language Models (LLMs) powered by Generative Artificial Intelligence (GenAI) offer a more integrated solution. However, domain-specific challenges persist, which led to the development of the Retrieval-Augmented Generation (RAG) framework. RAG enhances LLMs by incorporating external data retrieval, addressing issues related to data availability. To demonstrate RAG’s capabilities, we introduce the Political-RAG system, designed to extract political event information from media content, including Twitter data and news articles. Initially developed for event extraction, the Political-RAG system lays the foundation for developing various complex Political IE tasks. These include detecting hate speech, analyzing conflicts, assessing political bias, and evaluating social trends, sentiment, and opinions.

Deep learning for hate speech detection: a comparative study

Deep learning for hate speech detection: a comparative study

It’s ‘Gut Feeling’ Mostly: Online Hate, Uncivil Comments and Content Moderation in Australian Sports Media

There is evidence that online hate speech is increasing significantly on social media pages related to sport, but less research on how sport and media organisations are managing it. This research explored the management of content moderation in Australian sport media through qualitative interviews (16) with social media and communications staff in Australian sport and media organisations. It found that content moderation, or the moderating and removal of comments under social media posts, happened mostly as an addition to content creation work. Strategies for dealing with online hate and incivility were mostly the same between media and sport organisations, interviewees used some automated filters, but mostly manually hid, deleted and blocked comments and users to ‘clean’ their spaces. Overall there was a lack of formal guidelines and policies to direct moderation. Instead the work of content moderation was reliant on the actions of individuals, who took it on with a significant level of personal responsibility, and developed individual coping mechanisms to deal with the work. With its focus on communication staff in sport and media organisations this research contributes a different and important perspective to the growing field of research of online hate in sport.

Data retention in the light of copyright infringements: Protecting the status quo or seeking a third way? (Case C-470/21 La Quadrature du Net II)

In La Quadrature du Net II, the EU Court of Justice addressed – once again – the issue of the application of national retention legislation providing for the generalized retention of telecommunications data. This time, unlike in the previous cases, the problem was examined not in relation to measures used in the areas of combating serious crime or pursuing national security objectives, but in relation to counteracting online copyright infringements. Hence the case, decided by the CJEU sitting as the Full Court, concerned an important and topical issue, namely the possibility of countering mass copyright infringements in a situation where, according to a strict reading of the previous case law, the permissibility of imposing an obligation on telecommunications operators to collect information on users’ IP addresses was questioned. The LQDN II judgment – commented on by some as diverging from previous case law – is an important and necessary complement to the Court's earlier interpretation and, at the same time, a guideline on how to design retention laws to be applied not only in the area of countering copyright infringements, but also in other important areas, such as the fight against hate speech.

REGULAÇÃO LEGAL DAS REDES SOCIAIS: PROTEÇÃO DE DADOS E LIBERDADE DE EXPRESSÃO

The legal regulation of social networks is a field in constant evolution, which seeks to balance two fundamental pillars in the digital era: data protection and freedom of expression. Social networks manage a huge amount of personal data, which makes the existence of legal frameworks that protect users’ privacy imperative. Regulations such as the General Data Protection Regulation (GDPR) in the European Union and the General Data Protection Law (LGPD) in Brazil establish strict standards on how platforms must manage personal information. These laws not only allow users to access, correct or delete their data, but also impose sanctions on companies that do not comply with these requirements, ensuring a high level of security and transparency. At the same time, freedom of expression is a fundamental right that must be protected in the digital environment. However, exercising this right on social media often conflicts with the need to moderate content that may be considered offensive, dangerous or even illegal. Platforms implement moderation policies to mitigate the spread of hate speech, misinformation, and other harmful content. This, however, creates significant challenges in relation to censorship and the delimitation of permitted speech, which can affect the free expression of ideas and opinions. The challenge lies in finding an appropriate balance between protecting users’ privacy and data and enabling robust freedom of expression. An excessive focus on data protection can limit users’ ability to express themselves freely, while unrestricted freedom of expression can put individuals’ security and privacy at risk. As social media becomes more deeply integrated into everyday life and becomes a key platform for communication and the exchange of ideas, it is essential that regulations continue to adapt to address these dilemmas in a balanced and fair way. The future of regulation in this field will depend on the ability of legislators and platforms to adapt to rapid technological and social changes. ensuring that both rights are protected effectively and equitably.

‘Argumentative Disobedience’ as a Strategy to Confront Hate Speech

AbstractIn this paper, I examine argumentative strategies that social movements can follow to counter hate speech. I begin by reconstructing the disagreement space of the abortion debate in Argentina as a polylogue, identifying the protests of the social movement Pañuelos verdes as argumentative contributions. I then describe two different forms of hate speech used in response to the movement’s protests. I argue that hate speech discredits the position of Pañuelos verdes in the abortion debate and depicts their protests as social threats. Subsequently, I discuss three argumentative strategies that social movements can implement to address hate speech: arguing with hate speakers; advocating for a dialogue with restrictions; and opting for argumentative disobedience. Arguing with hate speakers aims to make hate speakers retract hate speech by exposing the undesirability of using hateful messages in argumentative exchanges. Advocating for a dialogue with restrictions aims to impose limited bans on public speech in order to ensure equal participation of arguers in argumentation. Finally, I propose the notion of argumentative disobedience to describe communicative responses to hate speech that aim to bring bystanders in line with the position of social movements in public debates.

In pursuit of ignorance: The institutional assault on disinformation and hate speech research

This article examines ongoing efforts to discourage disinformation and hate speech research in the U.S. These initiatives originate from a variety of institutional actors, including digital platforms, congressional committees, state attorneys general, and advocacy organizations. It employs the conceptual framework of agnotology, which refers to the study of deliberate culturally or politically induced ignorance. Through an analysis of over 1,800 pages of primary documents, it identifies the strategic approaches employed, the outcomes, and the broader democratic implications associated with these efforts to create and maintain knowledge gaps in the disinformation and hate speech arenas.

University Students’ Perceptions and Reactions to Cyberbullying on Social Media: The Case of Türkiye

ABSTRACT With cyberbullying emerging as a significant concern in contemporary society, understanding its prevalence, forms, and impacts is crucial for developing effective intervention strategies. This study aims to delve into the underlying factors contributing to cyberbullying incidents among university students. Employing quantitative methodologies, this research utilized a questionnaire comprising multiple-choice, scaled responses, and open-ended questions to explore students’ encounters, perceptions, and responses to cyberbullying. The study’s development was informed by theoretical frameworks and existing literature, with data analysis conducted using the Statistical Package for the Social Sciences (SPSS). The sample comprised 392 participants, predominantly female (79.1 percent) and Turkish (98.7 percent). Most participants were 18–21 (46.9 percent) and 21–25 (41.3 percent). The majority (84.9 percent) were studying at a state university. The study reveals a widespread acknowledgment of cyberbullying as a severe problem among social media users, particularly prevalent on platforms like Instagram and TikTok. Offensive comments and hate speech were identified as dominant forms of cyberbullying, alongside newer tactics like emoji-based ridicule. Participant responses ranged from passive bystander behavior to active intervention, reflecting diverse coping mechanisms and strategies. The findings underscore the need for comprehensive approaches to combat cyberbullying, including educational programs, institutional support mechanisms, and interventions promoting digital citizenship. Furthermore, fostering peer support networks and enhancing awareness in university settings is crucial for effectively addressing cyberbullying. The study highlights various intervention areas and approaches to mitigate the impact of cyberbullying and promote a healthier digital environment. Future research should continue exploring the underlying motivations and barriers to seeking help, aiming to develop targeted strategies for prevention and intervention.

Research on multimodal hate speech detection based on self-attention mechanism feature fusion

Research on multimodal hate speech detection based on self-attention mechanism feature fusion

Drivers of hate speech in political conversations on Twitter: the case of the 2022 Italian general election

We study the prevalence and characteristics of toxic speech on Twitter in the run-up to the 2022 Italian General election. We analyzed over 8.5 M tweets shared by 450k unique users and employed a machine learning classifier to estimate the toxicity score of their messages. We found that supporters of different political coalitions exhibit varying levels of toxic speech, sending between 6-8% of toxic messages overall. Notably, Centre-Left politicians received more toxic messages on average, with the largest target of hate receiving over 15,000 toxic replies. We employed Generalized Linear Models to study factors that drive hate speech to political targets, finding that, importantly, politicians employing more abusive and harmful language are also more likely to attract more inflammatory speech. Our findings underscore the critical need for targeted interventions to address hate speech online, fostering healthier dialogue and safeguarding democratic discourse.

GASCOM: Graph-based Attentive Semantic Context Modeling for Online Conversation Understanding

Online conversation understanding is an important yet challenging NLP problem which has many useful applications (e.g., hate speech detection). However, online conversations typically unfold over a series of posts and replies to those posts, forming a tree structure within which individual posts may refer to semantic context from elsewhere in the tree. Such semantic cross-referencing makes it difficult to understand a single post by itself; yet considering the entire conversation tree is not only difficult to scale but can also be misleading as a single conversation may have several distinct threads or points, not all of which are relevant to the post being considered. In this paper, we propose a Graph-based Attentive Semantic COntext Modeling (GASCOM) framework for online conversation understanding. Specifically, we design two novel algorithms that utilize both the graph structure of the online conversation as well as the semantic information from individual posts for retrieving relevant context nodes from the whole conversation. We further design a token-level multi-head graph attention mechanism to pay different attentions to different tokens from different selected context utterances for fine-grained conversation context modelling. Using this semantic conversational context, we re-examine two well-studied problems: polarity prediction and hate speech detection. Our proposed framework significantly outperforms state-of-the-art methods on both tasks, improving macro-F1 scores by 4.5% for polarity prediction and by 5% for hate speech detection. The GASCOM context weights also enhance interpretability.

From online hate speech to offline hate crime: the role of inflammatory language in forecasting violence against migrant and LGBT communities

Social media messages often provide insights into offline behaviors. Although hate speech proliferates rapidly across social media platforms, it is rarely recognized as a cybercrime, even when it may be linked to offline hate crimes that typically involve physical violence. This paper aims to anticipate violent acts by analyzing online hate speech (hatred, toxicity, and sentiment) and comparing it to offline hate crime. The dataset for this preregistered study included social media posts from X (previously called Twitter) and Facebook and internal police records of hate crimes reported in Spain between 2016 and 2018. After conducting preliminary data analysis to check the moderate temporal correlation, we used time series analysis to develop computational models (VAR, GLMNet, and XGBTree) to predict four time periods of these rare events on a daily and weekly basis. Forty-eight models were run to forecast two types of offline hate crimes, those against migrants and those against the LGBT community. The best model for migrant crime achieved an R2 of 64%, while that for LGBT crime reached 53%. According to the best ML models, the weekly aggregations outperformed the daily aggregations, the national models outperformed those geolocated in Madrid, and those about migration were more effective than those about LGBT people. Moreover, toxic language outperformed hatred and sentiment analysis, Facebook posts were better predictors than tweets, and in most cases, speech temporally preceded crime. Although we do not make any claims about causation, we conclude that online inflammatory language could be a leading indicator for detecting potential hate crimes acts and that these models can have practical applications for preventing these crimes.

Internalization of Religious Moderation Values Through Digital Literacy at Madrasah

The teachings of Rahmatan Lil Alamin Islam held by Muslims should also be instilled early, especially at the age of children. The social environment with diverse ethnicities, races, cultures and religions requires educators at the Madrasah Tsanawiyah (MTs) level to be present in providing moderate religious insights. Especially in the era of advances in communication technology, intolarance issues are also prone to be found on various social or digital media platforms. Therefore, religious moderations must be introduced and internalised to students. This article aims to explain the internalisation of religious moderation values through digital literacy at MTs PSM Randublatung Blora. This research is a descriptive qualitative field research, with data collection techniques of observation, interview, and documentation. The field data was analysed with descriptive qualitative method. The results of this study reveal that the internalisation of religious moderation is carried out with a class mentoring model and digital literacy training. Class mentoring to students as an effort to introduce and understand students related to religious moderation, followed by digital literacy training as an internalisation effort with students' ability to filter intolerant messages, and hate speech and be able to campaign with positive content containing religious moderation values.

Natural Language Processing for Sentiment Analysis in Social Media Posts to Identify Suspicious Behaviour

This article reports a study on the use of the natural language processing approach in the sentiment analysis of social media posts for detecting suspicious activities involving deep learning methods. The primary objective is to design an automated system capable of identifying unfavourable sentiments and patterns that imply suspicious activity such as cyberbullying, hate speech and disinformation. The methodology involves collecting a diverse data set of 2,000 posts from platforms such as X and Facebook, pre-processed by tokenisation and normalisation. The sentiment analysis model uses a long short-term memory (LSTM) network because it finds that the LSTM network is good for recognising long-term dependencies between texts. Other heuristics are used to give prominence to posts with malicious behavioural profiles to flag them accordingly. The model is designed and tuned to pre-labelled data sets. This work shows considerable improvement over previous techniques at categorising the posts as suspicious by showing higher accuracy, precision, recall and F1 score ratios. In doing so, this research positively impacts the enhancement of paranoid security or the totality of social media safety by offering a strong early detection mechanism for reducing risky behaviours on social networks.

Polarization and hate speech based on fuzzy logic and transformers: the case of the 2023 Spanish general elections

ABSTRACT Affective polarization in the digital debate of the Spanish presidential election campaign (2023), following the sudden call of the Spanish president on July 23, was measured. Using transformers, topics were detected, and sentiment analysis techniques were applied in the political debate during the elections to measure the emotional valence of the debate. The topics that dominate most of the debate are Candidates (n1 = 17170) and Opposition (n3 = 15327). These topics also show the highest typical polarization deviances. Based on affective polarization, a polarization measure (JDJ) grounded in the fuzzy sets was applied. The topic activism has the highest polarization value, while the topic of voting has the lowest. This analysis highlights a dichotomy that defines the Spanish political reality: the positive image of conventional political participation in the face of the rejection of collective action processes.

A Service-Learning Program to Prevent Online Hate Speech Perpetration Among Secondary Education Students: A Pilot Study.

The evolution and rapid expansion of online hate speech perpetration (OHSP) behaviors among adolescents toward different groups contrast with the lack of theory-based and empirically evaluated programs targeting this phenomenon. This study, developed by the Spanish Volunteering Platform, contributes to reversing this situation by investigating the impact of a service-learning program to prevent OHSP among students in secondary education. To do so, a cluster-randomized research design was adopted with a sample of secondary school students (n = 60) from two public schools. After the random assignment of each class group to either the intervention or control condition, the service-learning program previously designed was implemented for 3 months. Regression results in model 7 with all predictors and controls yield a significant effect on the variable experimental group = 0.803, p < .05, and gender = 0.987, p < .05, explaining 45.9% of the variance. The effect size for this intervention was 0.42. From a theoretical perspective, participants are moving in the right direction to reduce and/or prevent OHSP behaviors, which demonstrates the theoretical utility of the theory of planned behavior in hate speech research. Finally, several recommendations are made to improve the impact of future replications or adaptations of this program.

Public Perception of Hate Speech on Social Media (Facebook) during the 2023 General Elections in Nigeria

In the contemporary global landscape, the advent of information and communication technology has interconnected societies, fostering increased civic engagement. Notably, social media platforms, a manifestation of this technological innovation, play a vital role in political discourse, drawing active participation from politicians, political parties, and their supporters. However, the utilization of these platforms in the context of electoral campaigns, particularly in nations like Nigeria, has raised concerns due to the dissemination of political information carrying adverse repercussions. This study delves into the examination of the public's perceptions regarding hate speech on Facebook during Nigeria's 2023 general elections. Employing a survey research strategy and leveraging convenient sampling, data were collected from residents within the geographical scope of Kwali Area Council, Abuja, Nigeria. The research framework drew upon the public sphere theory to elucidate its core objectives. Findings underscored the prevalence of hate speech on Facebook during the 2023 elections, attributing it to factors such as financial inducements, conspiracies, lack of citizen journalist training, religious interests, and sectional, tribal, and political influences. Notably, Facebook emerged as a platform for disseminating speeches that fueled animosity toward politicians, tribes, religions, and sections, instigating widespread antipathy without substantial knowledge of the targeted individuals or candidates. The research also highlighted the use of Facebook for insults, harassment, and the propagation of divisive narratives, contributing to tribal and religious conflicts, heightened tensions, attacks, disparities, and internal discord. This study's implications suggest a potential undermining of Nigeria's democratic progress. As a remedy, the study advocates for the responsible and positive use of social media platforms by users, especially during critical events such as elections.

A Comparative Study of Transformer-based Models for Hate-Speech Detection in English-Kiswahili Code-Switched Social Media Text

The transformer architecture, first introduced in 2017 by researchers at Google, has revolutionized natural language processing in various tasks, including text classification. This architecture formed the basis of future models such as those used in hate speech detection in code-switched text. In this research, we conduct a comparative study of transformer-based models for hate speech detection in English-Kiswahili code-switched text. First, the models were compared as feature extractors using a traditional classifier and then as end-to-end classifiers. The three multilingual transformer-based models compared include mBERT, mDistilBERT and XLM-RoBERTa, using SVM as the traditional classifier for the extracted features. The HateSpeech_Kenya dataset, sourced from Kaggle, was utilized in this study. As a feature extractor, mBERT’s hidden states trained the highest-performing SVM with an accuracy of 0.5461 and a macro f1 score of 0.40. Among the three models evaluated, XLM-RoBERTa achieved the highest accuracy of 0.6069 and a macro f1 score of 0.49 on a balanced dataset. In contrast, mBERT achieved the highest accuracy of 0.7820 and a macro f1 score of 0.53 on an imbalanced dataset. The comparative study establishes that using transformer-based models as end-to-end classifiers generally performs better than using them as feature extractors with traditional classifiers. This is because directly training the models allows them to learn more task-specific features. Furthermore, the varying performance across balanced and imbalanced datasets highlights the need for careful model selection based on the dataset characteristics and specific task requirements.