Open research questions in Hate Speech and Cyberbullying Detection
50 unresolved questions extracted from the limitations and future-work sections of 668 Hate Speech and Cyberbullying Detection papers in our library. Each links back to the study that raised it.
What the literature leaves open
Reducing bias and discrimination in AI tools is not only morally required but also necessary to ensure that these widely used tools will not distort public opinion. The following recommendations are intended to ensure that researchers of all races, genders, and religions can use various AI tools fairly and responsibly to support all victims, whether within or outside the Middle East and North Africa. 1. Updating Google Gemini: Gemini's training data should be updated, and the ethical issues in it should be addressed to ensure that the widely used tool fairly represents all sides of any conflict, whether within or outside the Middle East and North Africa. 2. Bias detection: More detailed experiments should be performed on the training data of all the AI tools to ensure that there are no underlying agenda-promoting discriminatory outcomes 3. Human Element Involvement: Human insights should be considered, and any AI-generated content should be reviewed and vetted by human reviewers to ensure fair representation. 4. Explainability: AI model development whose content can be explained, making it transparent and understandable to the user for learning and understanding of bias. 5. Deployment: No machine learning algorithm should be deployed until the training data are reviewed, and given their size, the data should not drive any agenda. It runs on ethical principles, and these can be maintained through systems for detecting bias.
Algorithmic Framing of Conflict: A Comparative Analysis of Google Gemini and ChatGPT in the Israeli–Palestinian Context · 2026 · DOIVeriSphere, an intelligent cyberbullying detection system, was developed in this study. VeriSphere consists of a combination of NLP methods with XGBoost classifier for early detection of abusive contents on social networks. The proposed system consists of comprehensive text preprocessing, TF–IDF feature representation, and ensembles to precisely differentiate between bullying and non-bullying content with good predictive performance and high computation efficiency. VeriSphere achieved 92.84% accuracy, 91.76% precision, 92.31% recall, 92.03% F1-score, and 93.90% ROC–AUC which clearly shows the effectiveness and efficiency of the proposed system. In addition to this, the inclusion of an automated email alert mechanism increases the practicality of the system for timely moderation of online toxic behavior without putting much manual effort into it. Thus, it can be said that the proposed system provides an effective and reliable solution for real-time cyberbullying detection on social media platforms. The future research efforts are aimed at extending the framework for supporting multiple languages and code-mixed text using transformer-based language models for better context comprehension, multimodal cyberbullying detection using visual text, and developing adaptive learning algorithms that identify evolving cyberbullying patterns with computational efficiency for large scale real-time deployment.
VeriSphere: An NLP and XGBoost-Based Framework for Early Detection of Cyberbullying in Social Media · 2026 · DOICombining large-scale AI incident analysis with an intersectional rubric and thematic analysis, we surface both established and underexplored harm patterns.
Why AI Harms Can't Be Fixed One Identity at a Time: What 5300 Incident Reports Reveal About Intersectionality · 2026 · DOIFuture research should address these limitations by refining preprocessing pipelines, applying more rigorous stopword customization and integrating time-series analysis to examine how discourse on gender-based violence evolves in response to policy changes or high-profile cases. Second, the preprocessing stage was insufficient in filtering out all irrelevant and semantically ambiguous words, which contributed to the emergence of topics that were difficult to interpret in some modeling phases.
Sarcasm is an important aspect of any language. It includes expressing ideas, opinions and emotions in an indirect im- plicit way. This nature of implicitness makes sarcasm prob- lematic for SA systems which mostly rely on the surface meaning/features. In this work, we presented ArSarcasm, a new Arabic sar- casm dataset. The dataset was created through the re- annotation of available Arabic sentiment datasets. The new dataset contains sarcasm, sentiment and dialect labels. Analysis shows that sarcasm is highly prominent in senti- ment datasets with 16% of them being sarcastic. We also show the high subjective nature of such datasets, which was demonstrated by the change in sentiment labels in the new annotation. The experiments show the gap between SA systems’ performance on non-sarcastic tweets compared to sarcastic tweets, which urges the need to study such phe- nomena. Finally, our initial experiments on sarcasm detec- tion show that it is a challenging task. We believe that this dataset is a starting point in the di- rection of full study of sarcasm and figurative language in Arabic. However, due to the highly subjective nature of sar- casm, its reliance on world knowledge, cultural background and the perspectives of the communication parties, we be- lieve that the data collection procedure should incorporate more signals about these information. In the future, we hope to prepare a new dataset that incorporates more textual information. We also hope to study and analyse the differ- ences and similarities among sarcastic expressions used by Arabic speakers in different countries.
Future work will explore transformer-based models (BERT, RoBERTa) for context-aware classification, multimodal analysis incorporating image and audio content, graph-based user network analysis for behavioral pattern detection, and federated learning approaches for privacy-preserving model training on distributed social media data.
4, we will present and discuss the four most prevalent open issues (I1-I4) in research as our third contribution in Sec. 1 Open Issues in Multi-label Hate Speech Classification As regards training datasets for textual multi-label hate speech classification, one central issue is related to class imbalance (I1) [6, 48, 66, 89].
A Survey of Machine Learning Models and Datasets for the Multi-label Classification of Textual Hate Speech in English · 2026 · DOIOur results are based on the Hateful Memes benchmark and may not generalize across domains or languages. Weakly supervised CoT distillation from commercial MLLMs can introduce bias and un- faithful reasoning. GRPO increases computational cost and reliance on closed models may affect reproducibility. Explanation quality is assessed using automatic metrics that only partially reflect human judgment.
Our study is limited to a small set of countries, which may not fully capture global cultural and linguistic diversity, re- ducing the generalisability of our findings. Additionally, we used five LLM architectures, limiting insight into how geo- graphic bias varies across other models. While our proposed approach improves consistency, it may reduce sensitivity to context-specific signals. We limited training to one epoch and monitored the loss to reduce overfitting. The debiasing approach was tested solely on hate speech detection, leav- ing its efficacy for other tasks unexamined. Potential societal impacts include reinforcing bias if debiasing is ineffective or misused. To mitigate misuse, we advocate for transparency in model training and deployment and stress the importance of ongoing evaluation across diverse contexts.
Personalisation or Prejudice? Addressing Geographic Bias in Hate Speech Detection Using Debias Tuning In Large Language Models · 2026 · DOIWhile this research provides a robust foundation, the rapidly evolving digital landscape offers several avenues for further exploration: Transition to Vision Transformers (ViT): Future iterations could replace the CNN with Vision Transformers to better capture global dependencies in images, potentially identifying even more subtle bullying cues. Video-Based Detection: Expanding the model (e.g., TikTok/Reels) by incorporating Recurrent Neural Networks (RNNs) or 3D-CNNs to analyze temporal actions. OCR Integration: Implementing Optical Character Recognition (OCR) to extract and analyze text embedded inside images (memes), which is a common loophole for current filters. Decentralized AI Nodes: Exploring the deployment of this model as a lightweight "edge" node in Web 3.0 environments to enable privacy-preserving, localized moderation without central data storage. Multilingual and Cultural Adaptation: Training the NLP stream on diverse dialects and "slang" datasets to ensure the model is effective across different global demographics.
Hybrid Text–Image Fusion Model for Early Detection of Cyberbullying in Online Social Ecosystems · 2026 · DOICurrent work lacks exploration of adversarial misspelling generation and targeted attacks on NLP systems; there is no systematic study of which spelling error patterns are most harmful to different downstream tasks or how to generate minimal misspellings that degrade performance.
Few studies explicitly investigate the interaction between misspelling robustness and pre-trained language model size or architecture type (e.g., BERT vs. RoBERTa vs. GPT variants); the relationship between model capacity and resilience to spelling perturbations requires systematic investigation.
The survey shows inconsistency in how different studies introduce and evaluate misspellings—some use synthetic error injection, others use naturally occurring errors from social media or user logs; there is no standardized benchmark or shared evaluation framework for assessing misspelling robustness across methods.
Existing misspelling datasets focus on character-level edits (substitution, deletion, insertion) but do not systematically cover real-world error patterns such as homophone confusion, phonetic misspellings, visual character confusion, or intentional stylistic variations in social media text.
Prior work on misspellings in NLP has predominantly targeted English and a handful of European languages; robustness evaluation for misspelling handling in morphologically rich languages, low-resource languages, and non-Latin scripts remains largely unexplored.
The comparison table reveals that most misspelling handling approaches focus on only 1-2 downstream NLP tasks; there is a gap in evaluating misspelling robustness across diverse applications including machine translation, question answering, semantic similarity, named entity recognition, and sentiment analysis simultaneously.
The paper identifies the paradox of anonymous intimacy (freedom and vulnerability trade-offs including deception risk and diminished real-world social skills) but does not specify measurement instruments, longitudinal study designs, or control group comparisons needed to quantify these outcomes in users of different anonymous platforms or AI-driven companionship systems.
The paper proposes a constellations model explaining how individuals shift between dissociated intrapsychic states across different online environments, affecting disinhibition and personality expression. However, it lacks specific neuroimaging, longitudinal behavioural tracking, or computational modeling approaches to validate this dissociation mechanism in anonymous versus identified online communication contexts.
The authors recommend that future studies define the online social setting carefully to evaluate anonymity effects alongside other online situational variables (invisibility, eye-contact absence), yet they do not specify which experimental conditions, platform architectures, or user interaction scenarios should be systematically varied to isolate these factors in controlled studies.
The paper discusses AI-based sentiment analysis and emotion recognition chatbots for crisis support in anonymous environments but does not address how these tools should be adapted specifically for detecting and intervening in toxic online behaviour or cyberbullying within forum posts where multiple attackers and victims interact anonymously.
The paper advocates for pseudonymous reputation systems that balance anonymity with accountability in anonymous platforms. However, it does not specify how such systems should be evaluated in terms of preventing catfishing, deception, or toxic behaviour, nor does it detail the computational requirements or algorithmic approaches needed to track reputation while maintaining cryptographic anonymity.
The authors propose that online sense of unidentifiability should be reconceptualized as spanning three major factors (non-disclosure, invisibility, absence of eye-contact) rather than a binary state. However, they do not provide empirical measurement scales or operationalization methods to quantify these three dimensions independently in cyberbullying detection systems or forum post analysis.
The study found that eye-contact had a stronger main effect on negative online disinhibition than anonymity itself, contradicting conventional assumptions. However, the paper does not specify how eye-contact enforcement mechanisms should be technically implemented or tested across different anonymous platform designs (forums, chat applications, video-based systems) to validate this finding's generalizability.
Detection of machine-generated text is a key countermeasure for reducing the abuse of NLG models, and presents significant technical challenges and numerous open problems.
Despite being an underreported topic in the news media, gender-based violence (GBV) undermines the health, dignity, security and autonomy of its victims.
Big Data Techniques to Study the Impact of Gender-Based Violence in the Spanish News Media · 2023 · DOI
Most-cited papers in Hate Speech and Cyberbullying Detection
- Fanning the Flames of Hate: Social Media and Hate Crime · Journal of the European Economic Association · 2020 · 316 citations
- The superstar social media influencer: Exploiting linguistic style and emotional contagion over content? · Journal of Business Research · 2020 · 229 citations
- Hate Speech Epidemic. The Dynamic Effects of Derogatory Language on Intergroup Relations and Political Radicalization · Political Psychology · 2020 · 224 citations
- The Distorting Prism of Social Media: How Self-Selection and Exposure to Incivility Fuel Online Comment Toxicity · Journal of Communication · 2021 · 164 citations
- What is hate speech? Part 1: The Myth of Hate · Law and Philosophy · 2017 · 156 citations
- “Hey Everyone, I’m Drunk.” An Evaluation of Drinking-Related Twitter Chatter · Journal of Studies on Alcohol and Drugs · 2015 · 108 citations
- Persistent interaction patterns across social media platforms and over time · Nature · 2024 · 102 citations
- Hate Knows No Boundaries: Online Hate in Six Nations · Deviant Behavior · 2020 · 100 citations
- Thirty years of research into hate speech: topics of interest and their evolution · Scientometrics · 2020 · 90 citations
- “HOT” ChatGPT: The Promise of ChatGPT in Detecting and Discriminating Hateful, Offensive, and Toxic Comments on Social Media · ACM Transactions on the Web · 2024 · 78 citations
Most recent work
- “The Report Button is Just for Decoration”: · Journal of Online Trust and Safety · 2026
- Mapping the Italian Telegram ecosystem: communities, toxicity, and hate speech · Journal of Big Data · 2026
- The Relationship Between a CSAM Warning Messaging Chatbot and User Behavior on Pornhub · Victims & Offenders · 2026
- Analyzing User Descriptions of Child Sexual Abuse Material Posted on Darknet Forums: A Manual and Automated Content Analysis · Victims & Offenders · 2026
- Machine Understanding of Harms: Theory and Implementation · Knowledge · 2026
- A Review of AI and Deep Learning Approaches for Content Moderation · International Journal of Computer Applications Technology and Research · 2026
- Misspellings in natural language processing: A survey of recent literature · Natural language processing. · 2026
- The Psychology Behind Online Anonymity and Toxic Behaviour · International Journal for Research in Applied Science and Engineering Technology · 2026
- DETECTION OF CHILD PREDATORS CYBER HARASSERS ON SOCIAL MEDIA · International Journal of Engineering Research and Science & Technology · 2026
- An Explainable Hybrid Framework for Detecting Harmful Short Comments on Vietnamese Social Media Using TF-IDF and Lightweight Transformers · Journal of Science and Technology on Information security · 2026
Find a gap in your own Hate Speech and Cyberbullying Detection sub-topic
This page shows what the Hate Speech and Cyberbullying Detection literature already flags as unresolved. To narrow it to your specific question, run the guided finder — it searches the gap library on demand and checks candidates against 250M+ OpenAlex works.
Open the Research Gap Finder →