CoChat
Sign Up
Cited by 30Hate Speech and Cyberbullying DetectionSpam and Phishing DetectionMisinformation and Its Impacts

Cyberbullying detection: Exploring datasets, technologies, and approaches on social media platforms

Open access
Philipo, Adamu Gaston, Sarwatt, Doreen Sebastian, Ding, Jianguo, Daneshmand, Mahmoud, Ning, Huansheng·2026·ACM Computing Surveys
DOI 10.1145/3785654
Open this paper in CoChat — free accountOpen PDF

Abstract

Cyberbullying has become a major challenge in the digital era, and many people, especially adolescents, use social media platforms to communicate and share information. Some exploit these platforms to embarrass others through messages, e-mails, speech, and public posts, causing severe psychological harm to victims. This study reviews existing research on technologies, approaches, datasets, and evaluation metrics for cyberbullying detection, while highlighting future directions and key challenges. The findings show that traditional models work reasonably well with small datasets but require constant updates; machine learning models face feature extraction and linguistic limitations; deep learning models perform better but lack multilingual and cross-lingual capabilities; and large language models (LLMs) achieve the highest performance, offering flexibility and rich linguistic features but face issues of high-energy use and real-time applicability. Addressing technological, methodological, dataset, and linguistic challenges will improve cyberbullying detection, helping to protect online communication and promote social responsibility.

AI snapshot

Written from the paper’s full text.

TL;DR

This systematic review synthesizes 27 studies (2018–2023/2024) on cyberbullying detection across social media, comparing machine learning, deep learning, traditional, and large language model approaches. It finds supervised methods generally outperform unsupervised ones, with SVM, BiLSTM, and BERT highlighted as strong performers, while dataset imbalance, English-language bias, and real-time detection remain major challenges.

Key findings

  • Across 27 included studies, supervised methods outperformed unsupervised ones; SVM excelled among machine learning approaches, BiLSTM among deep learning approaches, and BERT among large language models. Traditional rule-based and character percentage matching also reported high performance (e.g., A=0.96, P=0.71, R=0.73, F=0.97 [112]).
  • Twitter was the most frequently used data source and English was predominant (74.1% of datasets in Figure 7); dataset issues included limited data, imbalance, inaccurate data, lack of generalization, and bias. Example: a Twitter sexist/racist/neither dataset had 10000:1 high class imbalance, and FormSpring had 6.1% cyberbullying vs. 93.9% non-cyberbullying instances.
  • Reported model performance varied: TF-IDF with Naive Bayes/SVM reached A=0.976, P=0.964, R=0.931, F=0.947 [88]; BERT-based models reached A=0.90, P=0.92, R=0.88, F=0.93 [113]; BiLSTM/CNN reached A=0.827, P=0.827, R=0.850, F=0.855 [102].
  • Key challenges were categorized as language-, dataset-, and approach-related: diverse linguistic structure, cultural variation, context misinterpretation, class imbalance, inconsistent annotations, limited model interpretability, and lack of real-time detection. Future directions include low-resource language datasets, automated annotation, multi-label classification, ensemble methods, LLMs, and cross-lingual approaches.

Methods

  • PRISMA-based systematic review; searched IEEE Xplore, SpringerLink, MDPI, ScienceDirect, ACM Digital Library, and other sources using cyberbullying-related keywords (e.g., harassment, hate, abuse, toxic, gender, sexism, racism) and language-resource terms.
  • Screening: 119 records of abstracts/titles; 50 irrelevant records excluded; 69 publications assessed after duplicates/irrelevant removed; ultimately 27 studies included. The review compared datasets, technologies, approaches, and evaluation metrics (accuracy, precision, recall, F1) across machine learning, deep learning, traditional, and LLM categories.

Limitations

  • Stated: the review may be incomplete because new forms of cyberbullying may emerge given dynamic online interactions.
  • Stated: public datasets often suffer from inaccuracies, imbalances, and limitations, impacting detection accuracy; open-source code availability and comparative studies are identified as challenges.
  • Evident from text: the review synthesizes only 27 studies and is heavily based on English/Twitter datasets, which may limit generalizability to low-resource languages and other platforms; no primary data collection or meta-analysis is reported.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3785654

Cites

133 of 133 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 98 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    The review identifies technical issues affecting deep learning performance, such as data dependency, generalization, interpretability, overfitting, dataset size, adaptability, and explainability.

    “The review also identifies several technical issues affecting performance, such as data dependency, generalization, interpretability, overfitting, dataset size, adaptability, and explainability.”
    Quote in the paperSubstantial claim
  • Findingnone given

    Traditional approaches, including rule-based and character percentage matching, effectively classify cyberbullying instances on social media.

    “The review shows that traditional approaches, including rule-based and character percentage matching, effectively classify cyberbullying instances on social media.”
    Quote in the paperNot a substantial claim
  • Findingnone given

    Statistical approaches like rule-based and character percentage matching techniques outperform current methods in cyberbullying detection on social media.

    “Analysis of traditional models reveals that statistical approaches like rule-based and character percentage matching techniques outperform current methods in cyberbullying detection on social media.”
    Quote in the paperNot a substantial claim
  • Findingown data

    More than one source has been used to collect cyberbullying datasets on social media platforms.

    “The review identified that more than one source had been used to collect cyberbullying datasets on social media platforms as follows:”
    Quote in the paperNot a substantial claim
  • Findingnone given

    The review identified that several studies used feature vector representation in both machine learning and deep learning models for cyberbullying text detection.

    “The review identified that several studies used feature vector representation in both machine learning and deep learning models for cyberbullying text detection.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 58
  • Classified as probably not a substantial claim. 66

  • May combine two claims.

    Findingown data

    Twitter is the most frequently used platform, and English is the predominant language in the datasets used in deep learning models for cyberbullying detection.

    “Twitter is the most frequently used platform, and English is the predominant language in the datasets.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Large language models lack real-time cyberbullying detection capability.

    “Large language models also lack real-time cyberbullying detection capability.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Reddit and YouTube datasets are used more frequently than those from other platforms, with English being the predominant language, in large language model cyberbullying detection.

    “Reddit and YouTube datasets are used more frequently than those from other platforms, with English being the predominant language.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Large language models, including GPT, BERT, XLNet, and XLM, effectively classify cyberbullying instances, especially BERT.

    “The review shows that large language models, including GPT, BERT, XLNet, and XLM, effectively classify cyberbullying instances, especially BERT.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    The review identified weaknesses of machine learning classifiers in cyberbullying detection, including lack of automatic feature extraction, inability to perform real-time detection, and difficulty processing large volumes of data.

    “The review identified weaknesses of machine learning classifiers in cyberbullying detection, including the lack of automatic feature extraction methods, inability to perform real-time detection, and difficulty processing large volumes of data [95,122,123].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [95]

      Not checked in the free run.

    • cites [122] source

      Not checked in the free run.

    • cites [123] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Findingnone given

    The review identifies technical issues affecting machine learning performance, such as data volume, time efficiency, false positives and negatives, real-time features, biases, context understanding, adaptability, dataset size, scalability, and interpretability.

    “The review also identifies several technical issues affecting performance, such as data volume, time efficiency, false positives and negatives, real-time features, biases, context understanding, adaptability, dataset size, scalability, and interpretability.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    Techniques such as TF-IDF, random, and GloVe are frequently used to enhance performance in deep learning cyberbullying detection.

    “Techniques such as TF-IDF, random, and GloVe are frequently used to enhance performance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The analysis of datasets used in deep learning models for cyberbullying detection identified challenges such as limited data usage, data imbalance, inaccurate data, lack of generalization, and dataset biases, affecting performance.

    “The analysis of datasets used in deep learning models for cyberbullying detection identified challenges such as limited data usage, data imbalance, inaccurate data, lack of generalization, and dataset biases, affecting performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Supervised deep learning methods outperform unsupervised ones in detecting cyberbullying on social media, with BiLSTM excelling.

    “Analysis of deep learning approaches reveals that supervised methods outperform unsupervised ones in detecting cyberbullying on social media, with the BiLSTM algorithm excelling.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    The analysis of datasets used in machine learning models for cyberbullying detection identified challenges such as limited data usage, data imbalance, inaccurate data, lack of generalization, and dataset biases, affecting performance.

    “The analysis of datasets used in machine learning models for cyberbullying detection identified challenges such as limited data usage, data imbalance, inaccurate data, lack of generalization, and dataset biases, affecting performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Analysis of datasets for large language models in cyberbullying detection reveals challenges such as limited data usage, data imbalance, lack of generalization, and real-time detection issues, affecting model performance.

    “Analysis of datasets for large language models in cyberbullying detection reveals challenges such as limited data usage, data imbalance, lack of generalization, and real-time detection issues, all of which affect model performance.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Twitter is the most frequently used platform, and English is the predominant language in the datasets used in machine learning models for cyberbullying detection.

    “Twitter is the most frequently used platform, and English is the predominant language in the datasets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingcitation

    Traditional classifiers struggle with automatic feature extraction, processing large data volumes, and understanding word meanings and semantics associated with cyberbullying.

    “Traditional classifiers struggle with automatic feature extraction, processing large data volumes, and understanding word meanings and semantics associated with cyberbullying [124].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [124] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    Techniques like TF-IDF and self-attention enhance performance in large language model cyberbullying detection.

    “Techniques like TF-IDF and self-attention enhance performance.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The results from the training and test phases demonstrate the effectiveness of various linguistic features in detecting cyberbullying.

    “The results from the training and test phases demonstrate the effectiveness of various linguistic features in detecting cyberbullying.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    TUKI and Ofcom were identified as data sources for Swahili and English, respectively, in traditional cyberbullying detection.

    “TUKI and Ofcom were identified as data sources for Swahili and English, respectively.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    Large language models are unsupervised approaches that excel in cyberbullying detection on social media, with BERT outperforming other algorithms.

    “Analysis of large language models reveals that these unsupervised approaches excel in cyberbullying detection on social media, with BERT outperforming other algorithms.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given

    Social media platform messages often feature informally written text that may not adhere to formal linguistic structures, making it challenging to distinguish patterns within the text.

    “Social media platform messages often feature informally written text that may not adhere to formal linguistic structures, making it challenging to distinguish patterns within the text.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Findingnone given

    The SVM algorithm and TF-IDF features are widely utilized in cyberbullying detection.

    “The SVM algorithm and TF-IDF features are widely utilized, while traditional approaches face limitations in processing large datasets and real-time detection.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    Existing public datasets for low-resource languages are limited in their coverage.

    “Moreover, there is a pressing need to create datasets specifically for these less powerful languages, as existing public datasets are limited in their coverage.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingreasoning

    Twitter is the primary dataset source in cyberbullying detection research, due to easy accessibility.

    “Twitter has emerged as the primary dataset source due to easy accessibility, with many datasets sourced from platforms like Facebook, Instagram, Reddit, Kaggle, etc.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingnone given

    Traditional cyberbullying detection approaches face limitations in processing large datasets and real-time detection.

    “The SVM algorithm and TF-IDF features are widely utilized, while traditional approaches face limitations in processing large datasets and real-time detection.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Findingnone given

    LLMs such as GPT and BERT play a significant role in enhancing cyberbullying detection accuracy.

    “LLMs such as GPT and BERT play a significant role in enhancing detection accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingnone given

    The review identified challenges in open-source code availability and comparative studies in cyberbullying detection.

    “The review has identified challenges in open-source code availability and comparative studies, emphasizing the need for resources in non-English languages.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Cyberbullying continues to occur due to limitations and weaknesses in the technologies and approaches used by existing detection algorithms.

    “However, cyberbullying continues to occur due to limitations and weaknesses in the technologies and approaches used by existing detection algorithms.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Cyberbullying strategies continuously develop and produce obfuscated words such as misspelling and coded languages that aim to harm individuals but are difficult for automatic detection systems to identify.
    • Rests on an assumption: Only Twitter platform allows researchers to clone unlabeled data via Twitter API for academic use without more restrictions compared to other social media platforms.
    • Rests on an assumption: Social media platforms are changing in terms of data availability, features, and user interfaces.
    • Rests on an assumption: Cyberbullying involves several forms such as multimodal and aggregation.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductcitation

    The systematic review employs the structured PRISMA framework to ensure both replicability and transparency.

    “The systematic review employs the structured PRISMA framework [12] to ensure both replicability and transparency.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [12] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Cyberbullying has profound negative effects on individuals and society, impacting them economically, socially, and psychologically, with adolescents highly vulnerable.

    “Cyberbullying has profound negative effects on individuals and society, impacting them economically, socially, and psychologically.Adolescents, in particular, are highly vulnerable to these effects [47,48].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [47]

      Not checked in the free run.

    • cites [48] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The increase of users on social media platforms is directly proportional to the increasing number of cyberbullying instances.

    “The increase of users on social media platforms is directly proportional to the increasing number of cyberbullying instances [4].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [4]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Most cyberbullying detection algorithms are developed for high-resource languages like English, Chinese, French, and German, causing bias in detecting low-resource languages like Swahili, Bengali, and Marathi, which may affect accuracy, precision, and fairness.

    “(2) Mostly of cyberbullying detection algorithms are developed to capture cyberbullying instances for high resource languages like English, Chinese, French, and Germany, so it causes bias in detecting cyberbullying instances on low resource languages like Swahili, Bengali, and Marathi which may affect the accuracy, precision and fairness of cyberbullying classification processes [7].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [7] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Deep learning methods accelerate learning and feature selection compared to traditional machine learning, enhancing real-time detection on social media platforms.

    “These methods accelerate learning and feature selection compared to traditional machine learning, enhancing real-time detection on social media platforms [95].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [95]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Deep learning approaches offer significant advantages in cyberbullying detection, efficiently managing large datasets, automatically extracting features, and classifying textual and visual instances.

    “Deep learning approaches offer significant advantages in cyberbullying detection, efficiently managing large datasets, automatically extracting features, and classifying textual and visual instances [94].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [94]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Sparse representations store only relevant values and their indices, enhancing efficiency.

    “Sparse representations store only relevant values and their indices, enhancing efficiency [121].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [121] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Cyberbullying reports have increased by nearly 70% despite social media platforms' algorithmic efforts.

    “Despite social media platforms' algorithmic efforts, cyberbullying reports have increased by nearly 70% [21].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The COVID-19 pandemic's shift to virtual interactions intensified cyberbullying instances, especially among adolescents aged 9 to 17.

    “The COVID-19 pandemic's shift to virtual interactions intensified cyberbullying instances, especially among adolescents aged 9 to 17 [18][19][20].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18][19][20]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Machine learning models require explicit feature engineering, while deep learning models can learn meaningful representations directly from raw data.

    “Machine learning models require explicit feature engineering, while deep learning models can learn meaningful representations directly from raw data [90].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [90] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Social media usage is increasing in developing countries, where people often communicate in their native languages, leading to numerous reported cyberbullying instances in these contexts.

    “Social media usage is increasing in developing countries, where people often communicate in their native languages, leading to numerous reported cyberbullying instances in these contexts [48].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [48] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Algorithms trained on past cases can classify new instances, reducing investigation time and resources.

    “Algorithms trained on past cases can classify new instances, reducing investigation time and resources [74].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [74]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Twitter provides a vast amount of data due to its large user base, resulting in a significant number of active users worldwide.

    “Moreover, Twitter provides a vast amount of data due to its large user base, resulting in a significant number of active users worldwide [9,131,132].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [9] source

      Not checked in the free run.

    • cites [131]

      Not checked in the free run.

    • cites [132] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Researchers may opt for Twitter due to its ease of data collection using APIs, real-time capabilities, and availability of a vast amount of textual data.

    “Researchers may opt for Twitter due to its ease of data collection using APIs, real-time capabilities, and the availability of a vast amount of textual data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Accurately detecting cyberbullying requires a comprehensive understanding of context and cultural intricacies, posing a challenge for algorithms.

    “Accurately detecting cyberbullying requires a comprehensive understanding of context and cultural intricacies, posing a challenge for algorithms.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Cyberbullying is often subjective and contingent on context.
    • Rests on an assumption: The interpretation of cyberbullying heavily relies on context.

  • May combine two claims.

    Premisecitation

    Low or accepted and moderate class imbalances are in a 10:1 ratio, but high-class imbalance ranges between 100:1 and 10000:1.

    “The low or accepted and moderate class imbalances are in a 10:1 ratio, but high-class imbalance ranges between 100:1 and 10000:1, as shown in Table 10 [146].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [146] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Disparities in cyberbullying can be attributed to cultural norms, educational systems, technological infrastructure, legal frameworks, and socio-economic disparities.

    “These disparities can be attributed to cultural norms, educational systems, technological infrastructure, legal frameworks, and socio-economic disparities [141].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [141] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Imbalance led models to become more motivated to detect the non-cyberbullying than the cyberbullying class and tended to label most instances as non-cyberbullying.

    “Firstly, bias in model training: this led models to become more motivated to detect the non-cyberbullying than the cyberbullying class and tended to label most instances as non-cyberbullying [108].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [108] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    Facebook and Instagram have stricter terms of service, data privacy policies, and ethical considerations regarding data use and fair use practices.

    “Additionally, these platforms have stricter terms of service, data privacy policies, and ethical considerations regarding data use and fair use practices.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Cyberbullying is a global phenomenon, with societal and cross-national differences in rates of bullying and victimization, as well as in the characteristics of cyberbullying instances.

    “Studies indicate that cyberbullying is a global phenomenon, with societal and cross-national differences in rates of bullying and victimization, as well as in the characteristics of cyberbullying instances [139,140].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [139]

      Not checked in the free run.

    • cites [140]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    A survey paper delves into the necessity for enhanced and more precise detection techniques, along with the exploration of alternative annotation methods.

    “A survey paper delves into the necessity for enhanced and more precise detection techniques, along with the exploration of alternative annotation methods [153].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [153] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Platforms differ in factors such as number of active users, restrictions on data collection, architectural design, and purpose.

    “Each of these platforms differs in various factors, such as the number of active users, restrictions on data collection, architectural design, and purpose.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning

    A deficiency in understanding the context of cyberbullying can impede advancements in research, prevention, and intervention activities.

    “In summary, a deficiency in understanding the context of cyberbullying can impede advancements in research, prevention, and intervention activities.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Annotation errors may be caused by subjective labeling, ambiguous definitions, cultural and contextual variations, inter-annotator agreement, and errors in the annotation process.

    “These errors may be caused by various reasons such as subjective labeling, ambiguous definitions, cultural and contextual variations, inter-annotator agreement, and errors in the annotation process [151].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [151]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The rise of social media platforms and communication in digital systems has introduced novel avenues for cyberbullying.

    “The rise of social media platforms and communication in digital systems has introduced novel avenues for cyberbullying [133].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [133] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Various cultural factors, including gender, race, ethnicity, disability, and sexual orientation, shape the dynamics of cyberbullying within workplace settings.

    “Various cultural factors, including gender, race, ethnicity, disability, and sexual orientation, shape the dynamics of cyberbullying within workplace settings [137].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [137] source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    In imbalanced classes, models predicting the non-cyberbullying class for every instance might still achieve high accuracy, but accuracy alone is insufficient for evaluating the true performance of the model.

    “Thirdly, misleading evaluation metrics: models predicting the non-cyberbullying class for every instance might still achieve high accuracy, but accuracy alone is insufficient for evaluating the true performance of the model in imbalanced classes.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Cyberbullying encompasses a variety of linguistic forms, including nonverbal modes of expression, code mixing, and the use of dissimilar languages.

    “Cyberbullying encompasses a variety of linguistic forms, including nonverbal modes of expression, code mixing, and the use of dissimilar languages.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisereasoning

    Facebook and Instagram are primarily designed for sharing photos and videos, leading to a higher prevalence of visual instances compared to text.

    “For instance, Facebook and Instagram are primarily designed for sharing photos and videos, leading to a higher prevalence of visual instances compared to text.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The imbalance of datasets is categorized into three groups: low-class, moderate-class, and high-class imbalance.

    “The imbalance of datasets is categorized into three groups: low-class, moderate-class, and high-class imbalance.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Datasets may be limited due to factors such as incomplete data percentages, inadequate details for comprehensive analytical evaluations, and the utilization of a small dataset for training learning models.

    “The dataset might be viewed as limited due to factors such as incomplete data percentages, inadequate details for comprehensive analytical evaluations, and the utilization of a small dataset for training learning models [120].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [120] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The Barlett Gentile cyberbullying model (BGCM) is a robust framework for predicting cyberbullying behaviors and attitudes across diverse cultural contexts.

    “The Barlett Gentile cyberbullying model (BGCM) has been identified as a robust framework for predicting cyberbullying behaviors and attitudes across diverse cultural contexts [138].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [138] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Automated methods like Self-Supervised Learning (SSL) models have been suggested for labeling data and enriching training sets using unlabeled data.

    “In response to this challenge, automated methods like Self-Supervised Learning (SSL) models have been suggested for labeling data and enriching training sets using unlabeled data [121].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [121] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A study introduces a labeled Instagram dataset with comprehensive annotations on critical cyberbullying characteristics.

    “Furthermore, a study introduces a labeled Instagram dataset with comprehensive annotations on critical cyberbullying characteristics, offering fresh perspectives for automated detection [152].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [152]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Detecting cyberbullying presents a challenge for multilingual pre-trained language models due to their struggle with extended social media platform sessions and the issue of imbalanced classes.

    “Detecting cyberbullying presents a challenge for multilingual pre-trained language models due to their struggle with extended social media platform sessions and the issue of imbalanced classes.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Transfer learning techniques show promising potential to address the challenge of low resources in many languages, especially low-resource languages like Swahili.

    “Due to the challenge of low resources in many languages, especially low-resource languages like Swahili, transfer learning techniques show promising potential to address this challenge.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Combining multiple detection techniques into an ensemble model can leverage the strengths of each approach while mitigating individual weaknesses, thereby improving overall detection accuracy.

    “Combining multiple detection techniques into an ensemble model can leverage the strengths of each approach while mitigating individual weaknesses, thereby improving overall detection accuracy.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    Expanding cyberbullying detection to various languages presents unique challenges specific to each language.

    “Expanding cyberbullying detection to various languages presents unique challenges specific to each language.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    Social media platforms encompass users from diverse regions who speak various languages.

    “Social media platforms encompass users from diverse regions who speak various languages.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Cross-lingual methods facilitate data augmentation by utilizing labeled data from one language to refine model training in languages with limited labeled data.

    “Furthermore, cross-lingual methods facilitate data augmentation by utilizing labeled data from one language to refine model training in languages with limited labeled data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    If data is of poor quality, showing behaviors like inconsistency, low quality, sparsity, noise, non-representativeness, ambiguous values, imbalance, and irrelevant features, resulting cyberbullying detection approaches may exhibit reduced accuracy.

    “Consequently, if the data is of poor quality, showing behaviors like inconsistency, low quality, sparsity, noise, non-representativeness, ambiguous values, imbalance, and irrelevant features, resulting cyberbullying detection approaches may exhibit reduced accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Current cyberbullying detection models are either incomplete or demonstrate restricted effectiveness when applied to real-time situations.

    “Current cyberbullying detection models are either incomplete or demonstrate restricted effectiveness when applied to real-time situations [157].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [157] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    By traversing input data in both directions, BiLSTMs potentially enhance training capabilities, leading to improved detection and prediction outcomes.

    “By traversing input data in both directions, BiLSTMs potentially enhance training capabilities, leading to improved detection and prediction outcomes [171].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [171]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Challenges of data can delay effective processing and lead to inaccurate conclusions, posing a significant challenge in deriving insights from the data.

    “Challenges of data can delay effective processing and lead to inaccurate conclusions, posing a significant challenge in deriving insights from the data.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Various forms of cyberbullying include sarcasm, toxicity, hate speech, and attacks based on ethnicity and religion, which are used to target individuals on social media platforms.

    “Additionally, there are various forms of cyberbullying, including sarcasm, toxicity, hate speech, and attacks based on ethnicity and religion, which are used to target individuals on social media platforms [22].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [22] source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Machine learning, deep learning, traditional, and large language models may detect cyberbullying instances on social media platforms, but they lack transparency in the decision-making process.

    “Both machine learning, deep learning, traditional, and large language models have shown that they may detect cyberbullying instances on social media platforms, but they lack transparency in the decision-making process.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Data augmentation enriches the training dataset, bolstering the generalization capability of detection models and heightening accuracy.

    “This augmentation enriches the training dataset, bolstering the generalization capability of detection models and heightening accuracy [168].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [168]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    GAN-based Oversampling (GBO) and Support Vector Machine-SMOTE-GAN (SSG) have been suggested to improve dataset balance and boost the performance of models.

    “In particular, novel approaches like GAN-based Oversampling (GBO) and Support Vector Machine-SMOTE-GAN (SSG) have been suggested to improve dataset balance and boost the performance of models [166].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [166]

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning

    Transfer learning and fine-tuning exhibit potential in cyberbullying detection but come with limitations and may not always represent the most optimal strategy.

    “Hence, while transfer learning and fine-tuning exhibit potential in cyberbullying detection, they come with limitations and may not always represent the most optimal strategy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationcitation

    Multilingual pre-trained language models provide a solution for applying models to low-resource languages without requiring additional training, thereby addressing the data availability gap between high and low-resource languages.

    “Multilingual pre-trained language models provide a solution for applying models to low-resource languages without requiring additional training, thereby addressing the data availability gap between high and lowresource languages.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Annotating data, such as categorizing, tagging, or labeling huge volumes of raw data, poses significant challenges for constructing discriminative supervised tasks.

    “Consequently, annotating data, such as categorizing, tagging, or labeling huge volumes of raw data, poses significant challenges for constructing discriminative supervised tasks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The bidirectional approach significantly improves the model's ability to understand dependencies along the sequence in both directions, making BiLSTM well-suited for tasks involving sequential data.

    “This bidirectional approach significantly improves the model's ability to understand dependencies along the sequence in both directions, making it well-suited for tasks involving sequential data [101].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [101]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Utilizing transfer learning and fine-tuning techniques demands extensive computational resources and time, along with abundant data to enhance the detection capabilities of models.

    “Utilizing transfer learning and fine-tuning techniques demands extensive computational resources and time, along with abundant data to enhance the detection capabilities of models [155].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [155] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    By training models on data from higher-resource languages like English and then adapting them for languages with lower resources, transfer learning increases the accuracy and performance in cyberbullying detection on social media platforms.

    “By training models on data from higherresource languages like English and then adapting them for languages with lower resources, transfer learning increases the accuracy and performance in cyberbullying detection on social media platforms.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The utilization of cutting-edge methods is constrained by restricted access to top-tier data, leading to the creation of small, varied datasets that impede straightforward comparisons of advancements.

    “The utilization of cutting-edge methods is constrained by restricted access to top-tier data, leading to the creation of small, varied datasets that impede straightforward comparisons of advancements [120].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [120] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    BiLSTM processes input sequences in both forward and backward directions, enabling it to capture information from both past and future contexts.

    “BiLSTM operates by processing input sequences in both forward and backward directions, enabling it to capture information from both past and future contexts.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    Data may contain more than one language in one instance, so code-switching modeling is needed to address this mixed-language challenge.

    “Sometimes, data may be found that contains more than one language in one instances, so code-switching modeling is needed to address this mixed-language challenge.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    These complexities involve cultural nuances, unconventional use of linguistic resources, and the availability of native language input methods.

    “These complexities involve cultural nuances, unconventional use of linguistic resources, and the availability of native language input methods [160].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [160] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    GANs hold potential for producing synthetic data to address dataset class imbalance challenges.

    “Generative Adversarial Networks (GANs) hold potential for producing synthetic data to address dataset class imbalance challenges [163][164][165].”
    Quote in the paperNot a substantial claim0 of 3 sources checked
    • cites [163]

      Not checked in the free run.

    • cites [164]

      Not checked in the free run.

    • cites [165]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Models like BERT and RoBERTa excel in natural language processing tasks but have not been broadly utilized for cyberbullying detection.

    “While models like BERT and RoBERTa excel in natural language processing tasks, they haven't been broadly utilized for cyberbullying detection [150].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [150] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    Certain characteristics revealing cyberbullying may be common across languages despite language differences.

    “Additionally, despite language differences, certain characteristics revealing cyberbullying may be common across languages.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    Users may express opinions, ideas, and comments in their respective languages, some of which may involve inappropriate language leading to online harassment.

    “However, these users may express opinions, ideas, and comments in their respective languages, some of which may involve inappropriate language leading to online harassment.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    Various classifiers may detect cyberbullying instances in different forms on social media platforms.

    “Various classifiers may detect cyberbullying instances in different forms on social media platforms.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Leveraging modern contextual language models such as BERT and slang-based word embeddings can greatly enhance cyberbullying detection, surpassing conventional models.

    “Furthermore, leveraging modern contextual language models such as BERT and word embeddings based on slang can greatly enhance the detection of cyberbullying, surpassing the capabilities of conventional models [177].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [177]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Using crowdsourcing techniques to replicate actual bullying instances in a controlled laboratory environment can produce credible data that augment existing datasets and boost classifier accuracy.

    “Using crowdsourcing techniques to replicate actual instances of bullying within a controlled laboratory environment can produce credible data, augmenting existing datasets and boosting the accuracy of classifiers [120].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [120] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The identification of cyberbullying instances relies on analyzing user data alongside personal characteristics such as name, gender, age, interests, and other human behaviors.

    “The identification of cyberbullying instances relies on analyzing user data alongside personal characteristics such as name, gender, age, interests, and other human behaviors.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    As detection algorithms advance, social media platforms must grapple with ethical issues related to privacy, security, bias, and censorship.

    “As detection algorithms advance, social media platforms must grapple with ethical issues related to privacy, security, bias, and censorship.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Improved detection could foster a more favorable online environment, especially for vulnerable demographics like youths and adolescents.

    “This, in turn, could foster a more favorable online environment, especially for vulnerable demographics like youths and adolescents.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationnone given

    Improving cyberbullying detection methods could enhance user safety by promptly recognizing and addressing harmful conduct.

    “Improving cyberbullying detection methods could enhance user safety by promptly recognizing and addressing harmful conduct.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Accurate detection of cyberbullying necessitates comprehension of various factors, including gender, religious affiliation, ethnicity, age, levels of aggression, and instances of non-cyberbullying.

    “Accurate detection of cyberbullying necessitates comprehension of various factors, including but not limited to gender, religious affiliation, ethnicity, age, levels of aggression, and instances of non-cyberbullying [182].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [182]

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Detecting cyberbullying in low-resource languages poses significant challenges due to linguistic intricacies, informal expressions, and the prevalence of slang.

    “Detecting cyberbullying in these low-resource languages poses significant challenges due to linguistic intricacies, informal expressions, and the prevalence of slang.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    By swiftly identifying and removing inappropriate or harmful instances without human intervention, automatic instances moderation helps social media platforms maintain a safe and positive online environment.

    “By swiftly identifying and removing inappropriate or harmful instances without human intervention, automatic instances moderation helps social media platforms maintain a safe and positive online environment.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    LLMs are adept at managing extensive datasets and continuously refining their performance through learning, thereby enhancing their efficacy over time.

    “Furthermore, LLMs are adept at managing extensive datasets and continuously refining their performance through learning, thereby enhancing their efficacy over time.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Striking a delicate balance between safety, user privacy, and the right to freedom of expression poses a complex challenge that requires careful deliberation.

    “Striking a delicate balance between safety, user privacy, and the right to freedom of expression poses a complex challenge that requires careful deliberation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The development of cross-lingual NLP models addresses the challenge of cross-linguistic and cross-cultural aspects of cyberbullying.

    “The development of cross-lingual NLP models addresses the challenge of cross-linguistic and cross-cultural aspects of cyberbullying.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    The development of multilingual frameworks simplifies the integration of data from multiple languages and cultural contexts to enhance performance in cyberbullying detection on social media platforms.

    “The development of multilingual frameworks simplifies the integration of data from multiple languages and cultural contexts to enhance performance in cyberbullying detection on social media platforms.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationreasoning

    Cyberbullying detection methods analyze user-generated instances to identify and remove material that violates community standards or legal requirements, such as offensive or abusive instances.

    “Cyberbullying detection methods analyze user-generated instances to identify and remove material that violates community standards or legal requirements, such as offensive or abusive instances.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Employing advanced models like the BiGRU-CNN sentiment classification model can notably improve cyberbullying detection by merging global and local features, attention mechanisms, and enhanced learning rates, resulting in superior classification accuracy.

    “Additionally, employing advanced models like the BiGRU-CNN sentiment classification model can notably improve cyberbullying detection by merging global and local features, attention mechanisms, and enhanced learning rates, ultimately resulting in superior classification accuracy [181].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [181]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given

    Cross-lingual NLP involves processing natural language data across diverse languages and simplifying the detection mechanism of cyberbullying instances on social media platforms.

    “This approach involves processing natural language data across diverse languages and simplifying the detection mechanism of cyberbullying instances on social media platforms.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Integrating textual, contextual, emotional, social attributes, and sentiment characteristics with conventional text features can elevate cyberbullying detection efficiency.

    “Integrating textual, contextual, emotional, social attributes, and sentiment characteristics together with conventional text features can elevate cyberbullying”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [73] source

      Not checked in the free run.

    • cites [130] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    Deep learning methods offer promise in handling large datasets and automatic feature extraction.

    “Deep learning methods offer promise in handling large datasets and automatic feature extraction.”
    Quote in the paperSubstantial claim
106 of 106

  • The paper relies on this without support.

    Cyberbullying strategies continuously develop and produce obfuscated words such as misspelling and coded languages that aim to harm individuals but are difficult for automatic detection systems to identify.

    “Apart from that there is continuous development of cyberbullying strategies which produce obfuscated words such as misspelling, and coded languages that aim to harm individuals but it is difficult to be identified by cyberbullying automatic detection systems, so, powerful tools are required to detect these kinds of cyberbullying instances.”

  • The paper relies on this without support.

    There is a significant number of users on social media platforms who communicate in English.

    “Additionally, there is a significant number of users on social media platforms who communicate in English.”

  • The paper relies on this without support.

    Translation tools integrated into social media platforms facilitate the translation of texts from other languages to English and vice versa.

    “Moreover, translation tools integrated into social media platforms facilitate the translation of texts from other languages to English and vice versa.”

  • The paper relies on this without support.

    Major social media platforms such as Twitter, Facebook, and Instagram have a large user base communicating primarily in English, providing ample English language data.

    “Major social media platforms such as Twitter, Facebook, and Instagram boast a large user base communicating primarily in English, thereby providing ample English language”

  • The paper relies on this without support.

    Cyberbullying occurs in any language.

    “The review reveals that most studies on cyberbullying detection focus on dominant languages like English, although cyberbullying occurs in any language.”

  • The paper relies on this without support.

    Only Twitter platform allows researchers to clone unlabeled data via Twitter API for academic use without more restrictions compared to other social media platforms.

    “Only Twitter platform allows researchers to clone unlabeled data via twitter API for academic use without more restrictions compared to other social media platforms with tight restrictions to get data from them.”

  • The paper relies on this without support.

    Social media platforms are changing in terms of data availability, features, and user interfaces.

    “(4) Social media platforms are changing in terms of data availability, features, and user interfaces.”

  • The paper relies on this without support.

    Cyberbullying involves several forms such as multimodal and aggregation.

    “Cyberbullying involves several forms such as multimodal and aggregation therefore, many cyberbullying detection algorithms are developed to classify cyberbullying instances in text based only which cannot classify all forms of cyberbullying.”

  • The paper relies on this without support.

    Cyberbullying is often subjective and contingent on context.

    “Cyberbullying is often subjective and contingent on context.”

  • The paper relies on this without support.

    The interpretation of cyberbullying heavily relies on context.

    “The interpretation of cyberbullying heavily relies on context.”

  • The paper relies on this without support.

    Labeling of datasets requires experts and linguists who have a comprehensive understanding of the language context to produce accurate datasets.

    “Labeling of datasets requires experts and linguists who have a comprehensive understanding of the language context to produce accurate datasets.”

  • The paper relies on this without support.

    Technology evolves rapidly, and bullies continuously develop new methods to harass others.

    “However, technology evolves rapidly, and bullies continuously develop new methods to harass others.”

  • The paper relies on this without support.

    Cyberbullying encompasses a range of behaviors, such as verbal abuse, threats, and the dissemination of inappropriate material.

    “However, cyberbullying encompasses a range of behaviors, such as verbal abuse, threats, and the dissemination of inappropriate material.”

  • The paper relies on this without support.

    Cyberbullying can take various linguistic forms, spanning text, images, and multimedia instances.

    “Moreover, cyberbullying can take various linguistic forms, spanning text, images, and multimedia instances.”

  • The paper relies on this without support.

    Deepfake instances are often used to harass or bully another person.

    “This instances is often used to harass or bully another person.”

  • The paper relies on this without support.

    Communities in sociology and psychology possess significant insights into cyberbullying, as they address various mental health issues stemming from online harassment.

    “Communities, especially those in sociology and psychology, possess significant insights into cyberbullying, as they address various mental health issues stemming from online harassment for various reasons.”
16 of 16

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Cyberbullying detection models and intervention tools are often trained around rigid and subjective cyberbullying criteria, potentially overlooking experiences that adolescents perceive as harmful.

    “This is particularly relevant given that cyberbullying detection models and intervention tools are often trained around the rigid and subjective cyberbullying criteria, potentially overlooking experiences that adolescents perceive as harmful (124, 125).”

    From A thematic analysis of adolescents’ perceptions of cyberbullying and cyberaggression: definitional distinctions and educational gaps (2026) · cites this paper as [125]

    Section 5.2.5 (page not provided) identifies subjective labeling and ambiguous definitions in datasets used to train detection systems; §5.2.7 says many studies use binary labels that miss other forms of cyberbullying. Section 5.2.1 warns that tools can miss the perspectives of those involved when context is misunderstood. The paper does not establish that intervention tools are often trained on rigid criteria or specifically that adolescents’ perceived harms are overlooked. (LLM, conf=0.84)

    Experimental — accuracy not yet measured

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.

  • Facebook imbalance ratio does not match its counts

    Minor

    The listed non-cyberbullying and cyberbullying counts give a ratio of about 15.06:1, not 10000:1.

    Reported 10000 · recomputed 15.062386

  • Social-media users are not 61% of the stated population

    Minor

    Dividing 4.95 billion users by a population of 7.888 billion gives about 62.75%, not 61%.

    Reported 61 · recomputed 62.75355

  • Table 6 F1 score exceeds its precision-recall upper bound

    Minor

    The traditional model's precision and recall average 0.72, whereas its reported F1 score is 0.97. An F1 score computed from those metrics cannot exceed their arithmetic mean.

    Reported 0.97 · recomputed 0.72

  • Table 2 F1 score exceeds its precision-recall upper bound

    Minor

    For the [48] row, the reported F1 score of 0.96 exceeds even the arithmetic mean of precision and recall, 0.87. An F1 score computed from those metrics cannot exceed that mean.

    Reported 0.96 · recomputed 0.87

  • Table 4 F1 score exceeds its precision-recall upper bound

    Minor

    For the [106] row, precision and recall average 0.83, yet the reported F1 score is 0.95. An F1 score computed from those metrics cannot exceed their arithmetic mean.

    Reported 0.95 · recomputed 0.83

  • Twitter class percentage does not match its counts

    Minor

    The three Twitter class counts total 15,054. The 10,000 instances classified as neither sexist nor racist therefore represent about 66.43%, not 68.4%.

    Reported 68.4 · recomputed 66.427528

  • Twitter offensive-class ratio does not match its counts

    Minor

    The offensive and non-offensive counts give a ratio of about 1.18:1, not 10:1.

    Reported 10 · recomputed 1.183718

  • Future directions are called three aspects but six are named

    Minor

    The sentence says the future directions comprise three aspects, then names six.

8 of 8

Cites

Flagged first, then by citations

Research works printed but not linked: 15

  1. Qais Faryadi. 2011. Cyber bullying and academic performance. International Journal Of Computational Engineering Research 1, 1 (2011), 2250–3005.
  2. Ujjwal Shrivastav, Prakash Srivastava, and Vikas Tripathi. 2023. A survey on cyberbullying detection techniques. In Proceedings of the 2023 10th International Conference on Computing for Sustainable Global Development (INDIACom). 628–633.
  3. 10.17148/ijarcce.2023.12681 · DOI
  4. Mr James Nambusi. 2018. Cyberbullying: Effect on work place production. Africa International Journal of Multidisciplinary Research 2, 1 (2018), 24–39.
  5. Hannah Beghin. 2020. The Effects of Cyberbullying on Students and Schools. BU Journal of Graduate Studies in Education 12 2 (2020) 19–22.
  6. Kris Varjas, Jasmaine Talley, Joel Meyers, Leandra Parris, and Hayley Cutts. 2010. High school students’ perceptions of motivations for cyberbullying: An exploratory study. Western Journal of Emergency Medicine 11, 3 (2010), 269–273.
  7. Chutikarn Saengmas Kulabutr Komenkul and Amporn Puapradit. 2018. Effects of social network and cyberbullying on mental health in young adults in Thailand. In Proceedings of the RSU International Research Conference 2018. Rangsit University. 454–460. DOI:10.14458/RSU.res.2018.105 · DOI
  8. Sumera Batool Rabia Yousaf and Feroza Batool. 2017. Bullying in social media: An effect study of cyber bullying on the youth. Pakistan Journal of Criminology 9 4 (2017) 119–139.
  9. Nazanin Alavi, Taras Reshetukha, Eric Prost, Kristen Antoniak, Charmy Patel, Saad Sajid, and Dianne Groll. 2017. Relationship between bullying and suicidal behaviour in youth presenting to the emergency department. J Can Acad Child Adolesc Psychiatry 26, 2 (2017), 70–77.
  10. Claudia Canali Michele Colajanni and Riccardo Lancellotti. 2011. Data Acquisition in Social Networks: Issues and Proposals. Retrieved from https://www.researchgate.net/publication/268360168. Accessed on 2024-05-03.
  11. Sweta Agrawal and Amit Awekar. 2018. Deep learning for detecting cyberbullying across multiple social media platforms. arXiv(2018), 1–12.
  12. M. Devi and M. Chitra Devi. 2018. Fuzzy based genetic operators for cyber bullying detection using social network data. International Journal of Scientific Research in Computer Science Engineering and Information Technology (IJSRCSEIT) 3 3 (2018) 437–444.
  13. Ke-Li Chiu, Annie Collins, and Rohan Alexander. 2021. Detecting hate speech with GPT-3. arXiv(2021), 1–29.
  14. Ulrik Brandes, Chandan Reddy, and Andrea Tagarelli. 2018. Are they our brothers? analysis and detection of religious hate speech in the arabic twittersphere. In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM).69–76.
  15. Mara Hamlett Grace Powell Yasin N. Silva and Deborah Hall. 2022. A Labeled Dataset for Investigating Cyberbullying Content Patterns in Instagram. Retrieved from www.aaai.org. Accessed on 2024-05-03.

Cited by

Sorted by citations

Related

Authors
Philipo, A. G., Sarwatt, D. S., Ding, J., Daneshmand, M., & Ning, H.
Affiliations
  • University of Science and Technology Beijing (CN)
  • Blekinge Institute of Technology (SE)
  • Stevens Institute of Technology (US)
ORCID
Venue
ACM Computing Surveys
Publisher
Association for Computing Machinery (ACM)
Published
2026-02-06
Volume
58
Issue
7
Pages
1-35
ISSN
  • 0360-0300 (ISSN-L)
  • 1557-7341
DOI
10.1145/3785654
OpenAlex
W4403749574
Semantic Scholar
55740b9b2c21cb7d10a5db65f0b60d8f71983265
Cited by
14
References
129
Citations by year
2026: 8 · 2025: 5 · 2024: 1
FWCI
5.20
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • ACM Computing Surveys (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
Funders
  • European Celtic+ and Swedish Vinnova (C2022/1-3)
  • Knowledge Foundation, Sweden, through the Human-Centered Intelligent Realities (20220068)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
  • Physical Sciences › Computer Science › Information Systems › Spam and Phishing Detection
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
Keywords
cyberbullying detection, social media, datasets, evaluation metrics, machine learning, deep learning, large language models, adolescents, online harassment, feature extraction

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-08.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-08.

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