The DOI matches its Crossref record; 197 of the 197 references it cites were found in a registry or an index; 64 of the 159 sources its claims cite were checked for support.
A Review on Deep-Learning-Based Cyberbullying Detection
Open accessAbstract
Bullying is described as an undesirable behavior by others that harms an individual physically, mentally, or socially. Cyberbullying is a virtual form (e.g., textual or image) of bullying or harassment, also known as online bullying. Cyberbullying detection is a pressing need in today’s world, as the prevalence of cyberbullying is continually growing, resulting in mental health issues. Conventional machine learning models were previously used to identify cyberbullying. However, current research demonstrates that deep learning surpasses traditional machine learning algorithms in identifying cyberbullying for several reasons, including handling extensive data, efficiently classifying text and images, extracting features automatically through hidden layers, and many others. This paper reviews the existing surveys and identifies the gaps in those studies. We also present a deep-learning-based defense ecosystem for cyberbullying detection, including data representation techniques and different deep-learning-based models and frameworks. We have critically analyzed the existing DL-based cyberbullying detection techniques and identified their significant contributions and the future research directions they have presented. We have also summarized the datasets being used, including the DL architecture being used and the tasks that are accomplished for each dataset. Finally, several challenges faced by the existing researchers and the open issues to be addressed in the future have been presented.
Overview
CoChat Summary
TL;DR
This review synthesizes deep-learning-based cyberbullying detection research, proposing a taxonomy/ecosystem covering data representation, DL models, and DL frameworks. It reports that DL generally outperforms traditional ML, with CNN/RNN-based and hybrid models showing strong results, while BERT is highlighted as a leading word-embedding approach.
Key findings
- The review selected 63 relevant primary articles after screening 1331 titles, covering English-language journals/conferences from January 2017 to January 2023.
- It reports that DL models outperform traditional ML for cyberbullying detection, and that CNN/RNN-based models (e.g., LSTM, Bi-LSTM) outperform other DL models such as GRU, DBN, MLPs, and BMs; hybrid models are described as more promising than individual models.
- BERT is described as the current state-of-the-art word embedding for text-based cyberbullying, and context-sensitive embeddings are said to perform better than context-insensitive ones; attention-based models (transformers, BERT, HAN, CNN-Att, LSTM-Att, GRU-Att) show strong results.
- Reported performance examples include Bi-GRU at 96.98% accuracy and 98.56% F1 on the Wikipedia Attack Dataset; Bi-LSTM at 96.5% accuracy and 98.69% F1 on the Wikipedia Web Toxicity Dataset; Bi-LSTM at 92.60% accuracy, 96.60% precision, and 94.20% F1 on Twitter; and CNN+Bi-GRU at 98.39% accuracy and 79.91% F1 on the Toxic Comment Classification Challenge.
Methods
- Review methodology: searched English-language journal and conference articles published January 2017–January 2023 in databases including IEEE Xplore, ScienceDirect, ACM Digital Library, Wiley, Springer Link, Taylor & Francis, and MDPI, plus patents, using Google Scholar keyword combinations such as “cyberbullying” and “deep learning.”
- Screening: started with 1331 article titles, removed duplicates, excluded low-tier journals/conferences, excluded articles not aligned with the research content, and excluded insignificant contributions, resulting in 63 selected primary research articles; an additional search used “deepfake” and “cyberbullying” in title, abstract, and keywords.
- Analysis: organizes the literature into a DL-based cyberbullying defense ecosystem/taxonomy covering data representation techniques, DL models, and DL frameworks; compares models, frameworks, datasets, and performance metrics, and discusses challenges and future trends.
Limitations
- The paper does not present a formal limitations section for the review itself. It states challenges in the field: lack of labeled/quality datasets, annotation requiring domain experts, cultural diversity, language/context difficulty, NLP ambiguity, data representation difficulty, limited reusability of sentiment-analysis models, and DL issues such as large data requirements, high computational power, unexplainable predictions, security attacks, and non-adaptivity.
- Evident from the text: the search is restricted to English-language articles and January 2017–January 2023, so it may miss other languages and periods; performance comparisons span heterogeneous datasets and metrics, so results are not directly comparable; and the review is qualitative/narrative with no meta-analysis or standardized quality appraisal described.
- The paper also notes a limited amount of cyberbullying datasets with images and text, and states that classification of cyberbullying in speech, videos, or deepfakes is hardly found.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
54 Valid · 14 Not valid · 42 Not checked · 67 Own results and interpretations
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“In our context, we found few cyberbullying detection works using TF-IDF [46,47].”
What the sources say
- Supportedcites [46] Riadi, 2019, Mobile Forensics for Cyberbullying Detection using Term Frequency-Inverse Document Frequency (TF-IDF), J. Eng. Sci. Technol. source
The Abstract and Research Method identify this paper as a cyberbullying-detection work using TF-IDF. The citing authors’ assessment that they found few such works is their own framing; this citation supports the paper as an example. (LLM, conf=0.90)
- Not checked — no full text of the sourcecites [47] Rahman, S., Talukder, K.H., and Mithila, S.K. (2021, January 14–16). An Empirical Study to Detect Cyberbullying with TF-IDF and Machine Learning Algorithms. Proceedings of the 2021 International Conference on Electronics, Communications and Information Technology (ICECIT), Khulna, Bangladesh. source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“DL models outperform MLmodels when we encounter complex linguistic expressions such as harassment with cyberbullying [25].”
What the sources say
- Partly supportedcites [25] Dadvar, M., and Eckert, K. (2018). Cyberbullying detection in social networks using deep learning based models; a reproducibility study. arXiv. sourceExperimental — accuracy not yet measured
The Conclusion (p. 11) reports that DNN models with transfer learning outperformed earlier ML results for cyberbullying detection on a YouTube dataset. It does not establish that the advantage occurs specifically with complex linguistic expressions or harassment, and the citing sentence generalizes beyond that dataset. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
The sources with a verdict support this claim; the rest weren’t checked yet, have no full text we could read, or their verdicts aren’t rated yet.
In the paper
“These studies do not show publicly available datasets for cyberbullying detection, and a few survey papers address the open issues and challenges [26,27].”
What the sources say
- Supportedcites [26] Salawu, 2017, Approaches to automated detection of cyberbullying: A survey, IEEE Trans. Affect. Comput. source
The Introduction identifies this paper as a survey of cyberbullying detection research. Section 4, “Research Challenges,” discusses key open issues, including the lack of a universally adopted definition and a shortage of large labelled corpora. (LLM, conf=0.93)
- Not checked — no full text of the sourcecites [27] Rosa, 2019, Automatic cyberbullying detection: A systematic review, Comput. Hum. Behav. source
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“Studies [22,58-60] show that BERT achieved high accuracy and F1 scores to classify cyberbullying in various types of online content, such as tweets and comments.”
What the sources say
- Not checked — no full text of the sourcecites [22] Paul, 2022, CyberBERT: BERT for cyberbullying identification, Multimed. Syst. source
- Not checked — no full text of the sourcecites [58] Mazari, A.C., Boudoukhani, N., and Djeffal, A. (2023). BERT-based ensemble learning for multi-aspect hate speech detection. Clust. Comput., 1–15. source
- Not matchedcites [59] Coban, O., Ozel, S.A., and Inan, A. (2023). ACM Transactions on Asian and Low-Resource Language Information Processing, Association for Computing Machinery.
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
- Partly supportedcites [60] Mozafari, M., Farahbakhsh, R., and Crespi, N. (2020, January 10–12). A BERT-based transfer learning approach for hate speech detection in online social media. Proceedings of the Complex Networks and Their Applications VIII: Volume 1 Proceedings of the Eighth International Conference on Complex Networks and Their Applications COMPLEX NETWORKS 2019, Lisbon, Portugal. sourceExperimental — accuracy not yet measured
The Results Analysis (p. 8) reports strong F1 scores for a BERT-based classifier on two Twitter hate-speech datasets. It reports precision, recall, and F1—not accuracy—and the study does not evaluate cyberbullying classification across content types such as comments. (LLM, conf=0.89)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
Every source the paper cites for this claim was checked and supports it.
In the paper
“The study found that the majority of language models achieved high F1 scores compared to traditional word embeddings such as BoW and TF-IDF, as well as semantic word embeddings such as word2vec, GloVe, and FastText.”
What the sources say
- Supportedcites [76] Pericherla, 2021, Performance analysis of word embeddings for cyberbullying detection, Proceedings of the IOP Conference Series: Materials Science and Engineering source
The Results and Discussion section states that most language models scored higher F1 scores than BoW, TF-IDF, word2vec, GloVe, and FastText in the study’s cyberbullying-detection experiments. This directly supports the attributed finding. (LLM, conf=0.95)
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
55 more claims · 18 with their cited sources checked
Create a free account to see all 177 claims and to verify the 4 references still open.
- It may state two separate claims as one. 62
- Probably not a substantial claim. 86
- The quote only partly states this claim. 2
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 5.9. Hybrid Models (LSTM-CNN, CNN-LSTM)
“Then, the main structure of this architecture is built with a Bidirectional LSTM layer followed by a CNN layer”
§ 5.9. Hybrid Models (LSTM-CNN, CNN-LSTM)
“LSTM-CNN performs better than CNN-LSTM because the CNN layer would receive the word embeddings as input, which will further be pooled to a smaller dimension, and then, the LSTM layer will use the ordering of said features to learn about the input's text ordering.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ 4.1.7. BERT
“Wikipedia (2500 million words) and Book Corpus (800 million words) are part of a vast corpus of unlabeled text that has been pre-trained.”
§ 4.1.7. BERT
“BERT is also pre-trained by using unlabeled data collected from the English Wikipedia and BooksCorpus, each of which contains 800 million words.”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Applicability of Different DL Frameworks
“We have also observed other research that examined cyberbullying by using different frameworks, such as Caffe [187], Deeplearning4j [188], and MXNet [178].”
§ 7. Deep Learning Frameworks
“| Caffe | • • | convolutional neural networks. Some pre-trained models are available. Since Python and MATLAB are very high-level languages, it is very easy to code with Caffe. | effectiveness day by day. • Many RNNapplications need variable-sized input because the static model graph does not fit. • Custom layers of Caffe must be written in C++ language. | Initially designed for CNNs [173] | No Works Found”
Citations
Cites
Hochreiter, S., & Schmidhuber, J. (1997) · Neural Computation · doi:10.1162/neco.1997.9.8.1735 · Cited by 102,318 · cited 2×
“…difficult to train, and unable to handle very long sequences when tanh or ReLU is used as the activation function Long Short-Term Memory (LSTM) [109] Social media content [7,21,68], Wikipedia, Twitter, Formspring and YouTube [25], CyberBERT [22], Bangla text [18], Indonesian language [64],…”
MethodBERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325
Glove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071 · cited 2×
“Pretrained word-embedding techniques include but are not limited to Word2Vec [32,33], GloVe [34], ELMo [35], fastText [36], and BERT [37] whereas One-hot encoding and TF-IDF [38] are nonpretrained.”
Method(2014) · Cited by 33,419
PyTorch: An Imperative Style, High-Performance Deep Learning Library
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.01703 · Cited by 31,473 · cited 1×
“Another widely used framework for cyberbullying detection [19,73,168,185] is PyTorch, which is developed based on the Torch [186] library of Lua programming language.”
MethodEfficient Estimation of Word Representations in Vector Space
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1301.3781 · Cited by 26,568
Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation
Cho, K., van Merrienboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1179 · Cited by 25,545 · cited 1×
“Deep Neural Networks such as Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU) [8], Long Short-term Memory (LSTM) [9], Bi-LSTM [10] and several other DL models can be used to detect this problem.”
MethodDistributed Representations of Words and Phrases and their Compositionality
Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1310.4546 · Cited by 24,533 · cited 2×
“Pretrained word-embedding techniques include but are not limited to Word2Vec [32,33], GloVe [34], ELMo [35], fastText [36], and BERT [37] whereas One-hot encoding and TF-IDF [38] are nonpretrained.”
MethodReducing the dimensionality of data with neural networks
Hinton, G. E., & Salakhutdinov, R. R. (2006) · Science · doi:10.1126/science.1127647 · Cited by 20,928
Research works printed but not linked: 17
- Feinberg, 2009, Cyberbullying, Educ. Dig., 74, 26
- Ramos, 2003, Using tf-idf to determine word relevance in document queries, Proceedings of the First Instructional Conference on Machine Learning, Volume 242, 29
- Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial nets. Adv. Neural Inf. Process. Syst., 27.
- Coban, O., Ozel, S.A., and Inan, A. (2023). ACM Transactions on Asian and Low-Resource Language Information Processing, Association for Computing Machinery.
- Kountchev, R., Rubin, S., Milanova, M., Todorov, V., and Kountcheva, R. (2008, January 29–31). Cognitive image representation based on spectrum pyramid decomposition. Proceedings of the WSEAS International Conference on Mathematical Methods and Computational Techniques in Electrical Engineering (MMACTEE), Athens, Greece.
- Sudirman, S., and Qiu, G. (2000). Colour image representation using BSP Tree. Proc. CGIP, 1–4. Available online: http://www.cs.nott.ac.uk/~pszqiu/Online/CGIP2000.pdf.
- Gao, 2021, Chinese cyberbullying detection with OCR and deep learning, J. Intell. Fuzzy Syst., 40, 4731
- Borah, N., and Borah, P. (2020, January 13–15). Detecting cyberbullying on social media using OCR and machine learning. Proceedings of the 4th International Conference on Intelligent Computing and Control Systems (ICICCS 2020), Madurai, India.
- Prechelt, L. (1998). Neural Networks: Tricks of the Trade, Springer.
- Das, H.S., and Roy, P. (2019). Intelligent Speech Signal Processing, Elsevier.
- Hinton, G.E. (2012). Neural networks: Tricks of the Trade, Springer.
- Goodfellow, I., Bengio, Y., and Courville, A. (2016). Deep Learning, MIT Press.
- Ritter, H., Martinetz, T., and Schulten, K. (1992). Neural Computation and Self-Organizing Maps: An Introduction, Addison-Wesley Reading.
- Vishwamitra, N., Hu, H., Luo, F., and Cheng, L. (2020, January 14–17). toward understanding and detecting cyberbullying in real-world images. Proceedings of the 2020 19th IEEE International Conference on Machine Learning and Applications (ICMLA), Online.
- Hodnett, M., and Wiley, J.F. (2018). R Deep Learning Essentials: A Step-by-Step Guide to Building Deep Learning Models Using TensorFlow, Keras, and MXNet, Packt Publishing Ltd.
- Shahane, 2019, Detection of Fake Profiles on Twitter using Random Forest & Deep Convolutional Neural Network, Int. J. Manag. Technol. Eng., 9, 3663
- Weischedel, R.M., and Bates, M. (2006). Challenges in Natural Language Processing, Cambridge University Press.
Cited by
Cyberbullying Sexism Harassment Identification by Metaheurustics-Tuned eXtreme Gradient Boosting
Dobrojevic, M., Jovanovic, L., Babic, L., Cajic, M., Zivkovic, T., Zivkovic, M., Muthusamy, S., Antonijevic, M., & Bacanin, N. (2024) · Computers, Materials & Continua · doi:10.32604/cmc.2024.054459 · Cited by 75
Detecting cyberbullying using deep learning techniques: a pre-trained glove and focal loss technique
Koshiry, A. M. E., Eliwa, E. H. I., El‐Hafeez, T. A., & Khairy, M. (2024) · PeerJ Computer Science · doi:10.7717/peerj-cs.1961 · Cited by 52
“While classical machine learning algorithms have been used to identify cyberbullying, deep learning techniques have emerged as a promising alternative ( Hasan et al., 2023 ).”
BackgroundFrom Text to Insight: An Integrated CNN-BiLSTM-GRU Model for Arabic Cyberbullying Detection
Daraghmi, E. Y., Qadan, S., Daraghmi, Y., Yousef, R. N. M., Cheikhrouhou, O., & Baz, M. (2024) · IEEE Access · doi:10.1109/access.2024.3431939 · Cited by 52
“Continued research in this domain has leveraged different Machine Learning (ML) and Deep Learning (DL) methodologies to make promising achievements in detecting and preventing text-based cyberbullying [2], [3], [4], [5], [6], [7], [8], [9], [10], [11].”
BackgroundAlqahtani, A. F., & Ilyas, M. (2024) · Machine Learning and Knowledge Extraction · doi:10.3390/make6010009 · Cited by 52
Detecting Cyberbullying using Machine Learning Approaches
Haider, A., Siddique, A. B., Ali, R. H., Imad, M., Ijaz, A. Z., Arshad, U., Ali, N., Saleem, M., & Shahzadi, N. (2023) · International Conference on IT and Industrial Technologies (ICIT) · doi:10.1109/icit59216.2023.10335773 · Cited by 35
“Then, in order to identify cyberbullying, these characteristics are utilized to train several machines learning algorithms, including Logistic Regression, Decision Tree, Random Forest, and SVM with kernel, and Naive Bayes [10].”
Method
Metadata
- Authors
- Hasan, M. T., Hossain, M. A. E., Mukta, M. S. H., Akter, A., Ahmed, M., & Islam, S.
- Affiliations
- United International University (BD)
- Edith Cowan University (AU)
- Corresponding author
- Islam, Salekul
- Venue
- Future Internet
- Publisher
- MDPI AG
- Published
- 2023-05-11
- Volume
- 15
- Issue
- 5
- Pages
- 179
- ISSN
- 1999-5903 (ISSN-L)
- OpenAlex
- W4376280098
- Cited by
- 103
- References
- 167
- Citations by year
- 2026: 30 · 2025: 37 · 2024: 32 · 2023: 4
- FWCI
- 9.71
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- Locations
- Future Internet (publishedVersion, cc-by)
- Australasian Journal of Paramedicine (submittedVersion, cc-by)
- RePEc: Research Papers in Economics (submittedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Institute for Advanced Research Publication Grant of United International University (IAR-2023-Pub-013)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
- Physical Sciences › Computer Science › Information Systems › Spam and Phishing Detection
- Social Sciences › Psychology › Social Psychology › Bullying, Victimization, and Aggression
- Keywords
- cyberbullying detection, deep learning, online harassment, text classification, deep learning architectures, data representation, mental health
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-08, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-08 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: A Review on Deep-Learning-Based Cyberbullying Detection. https://app.cochat.ai/papers/10.3390/fi15050179
- The paper
- Hasan, M. T., Hossain, M. A. E., Mukta, M. S. H., Akter, A., Ahmed, M., & Islam, S. (2023). A Review on Deep-Learning-Based Cyberbullying Detection. Future Internet. https://doi.org/10.3390/fi15050179
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.