Source-linked AI summary
Automatic Detection of Cyberbullying in Social Media Text
Cynthia Van Hee, Gilles Jacobs, Chris Emmery, Bart Desmet, Els Lefever, Ben Verhoeven, Guy De Pauw, Walter Daelemans, Véronique Hoste
TL;DR
Cyberbullying on social media creates a need to identify harmful messages automatically. This paper builds annotated English and Dutch data, models posts from bullies, victims, and bystanders with linear SVMs, and reports promising holdout performance that outperforms keyword and n-gram baselines.
Problem
Limited suitable public datasets make it difficult to develop models for automatic cyberbullying detection, despite the need to identify harmful social-media content.
Method
The paper creates fine-grained English and Dutch ASKfm annotations and trains linear SVM classifiers using varied feature types to model posts from bullies, victims, and bystanders.
Results
The optimised classifiers achieved F1-scores of 64.32% for English and 58.72% for Dutch, significantly outperforming keyword and unoptimised n-gram baselines.
Takeaways & Limitations
The experiments support automatic detection of cyberbullying signals across posts written by bullies, victims, and bystanders in English and Dutch social-media data.
Takeaways & Limitations
The approach depends on annotated data being available, while suitable public datasets for cyberbullying detection remain scarce.
Abstract
from arXiv · showhide
While social media offer great communication opportunities, they also increase the vulnerability of young people to threatening situations online. Recent studies report that cyberbullying constitutes a growing problem among youngsters. Successful prevention depends on the adequate detection of potentially harmful messages and the information overload on the Web requires intelligent systems to identify potential risks automatically. The focus of this paper is on automatic cyberbullying detection in social media text by modelling posts written by bullies, victims, and bystanders of online bullying. We describe the collection and fine-grained annotation of a training corpus for English and Dutch and perform a series of binary classification experiments to determine the feasibility of automatic cyberbullying detection. We make use of linear support vector machines exploiting a rich feature set and investigate which information sources contribute the most for this particular task. Experiments on a holdout test set reveal promising results for the detection of cyberbullying-related posts. After optimisation of the hyperparameters, the classifier yields an F1-score of 64% and 61% for English and Dutch respectively, and considerably outperforms baseline systems based on keywords and word unigrams.
1 Introduction
Cyberbullying is a serious and growing threat to young people, while online information overload makes comprehensive manual monitoring impractical. The paper therefore develops automatic detection of varied textual signals, including posts by bullies, victims, and bystanders.
- Motivation: Cyberbullying is increasingly prevalent among young people and can harm victims’ self-esteem, academic achievement, emotional well-being, and safety.Reported cybervictimisation and exposure rates vary across studies, partly because definitions, locations, and participant groups differ.
- Motivation: Successful early detection matters because moderators cannot manually monitor all user-generated content on the Web.The paper motivates intelligent systems that process information quickly and signal potential threats so moderators can respond before situations escalate.
- Research gap: Keyword-based moderation often misses implicit or subtle cyberbullying that lacks explicit harmful vocabulary.The paper identifies a need for intelligent, self-learning systems that go beyond keyword spotting to improve recall.
- Approach: The study detects a broader range of cyberbullying signals, including posts written by bullies, victims, and bystanders.This extends beyond approaches focused primarily on detecting cyberbullying attacks.
- Approach: The proposed approach uses a linear SVM with a varied feature set to detect cyberbullying automatically in English and Dutch social media text.The methodology is presented as language- and genre-independent when annotated data are available.
2 Related Research
Prior work established the feasibility of automatic cyberbullying detection but relied mainly on limited datasets and attacks by bullies. This paper broadens detection to multiple roles and forms using English and Dutch social-media corpora.
- Conceptual foundations: Cyberbullying definitions remain contested because power imbalance and repetition can be difficult to establish in online settings.Less strict definitions consequently do not always require these conditions.
- Detection approaches: Research on cyberbullying detection has used supervised and semi-supervised machine-learning approaches, commonly combining linguistic and sentiment features.These approaches typically learn from labeled data or combine labeled and unlabeled data.
- Data limitations: Suitable data remain a key bottleneck because few publicly available datasets support models that characterize cyberbullying.Existing studies therefore often use workshop datasets or construct corpora from platforms such as YouTube, Formspring, and ASKfm.
- Research scope: Earlier systems mainly focused on attacks or bully-authored posts, while coverage of other cyberbullying forms was unclear.The literature did not consistently establish whether sexual intimidation, harassment, or psychological threats were included alongside insults and derogatory language.
- Contribution: This study models cyberbullying as a broader phenomenon including attacks, victim reactions, and bystander reactions.Its detection system is intended to support manual monitoring on social networks.
- Experimental direction: The paper uses English and Dutch social-media data for binary classification with a rich feature set and optimized hyperparameters.The experiments evaluate the viability of detecting cyberbullying-related posts in both languages.
3 Data Collection and Annotation
The authors construct English and Dutch ASKfm corpora and annotate posts at both document and subsentence levels. The scheme captures cyberbullying roles and fine-grained harmful-language categories, although several categories show interpretive difficulty.
- Data collection: The study constructs English and Dutch ASKfm corpora containing social-media posts manually annotated for cyberbullying.ASKfm provides question-answer pairs posted on user profiles, including anonymously authored content.
- Categories: Fine-grained categories include threats, insults, discrimination, curses or exclusion, defamation, sexual talk, defense, and encouragement to the harasser.The scheme distinguishes subtypes such as general insults, attacks on relatives, racism, sexism, victim defense, and bystander defense.
- Roles: The annotation scheme assigns four roles: bully, victim, bystander-defender, and bystander-assistant.Passive bystanders are excluded because they are difficult to recognize in online text.
- Annotation procedure: Annotation operates at post level for cyberbullying relevance and author role, then at subsentence level for fine-grained cyberbullying spans.Annotators identify all spans corresponding to categories in the annotation scheme.
- Reliability: Trained linguists independently annotate both corpora, with agreement for cyberbullying identification of κ= 0.69 for Dutch and κ= 0.59 for English.Cohen’s Kappa is used for Dutch and Fleiss’ Kappa for English.
- Reliability: Defamation, curse, and encouragement categories are sometimes difficult to recognize and distinguish from insults or threats.Encouragements can depend on interpretation, especially for expressions such as “hahaha” or “LOL.”
4 Experimental Setup
The study evaluates binary detection of cyberbullying-related posts using linear SVMs, rich heterogeneous features, and optimised hyperparameters. English and Dutch experiments model attacks alongside victim and bystander reactions, with validation and hold-out testing.
- Task definition: Binary classification targets signals of cyberbullying, including posts by bullies, victims, and bystanders that may indicate an event.These signals are intended for further investigation by human moderators.
- Data: The English and Dutch corpora contain 113,698 and 78,387 posts, respectively, with heavily imbalanced class distributions.Cost-sensitive SVM weighting penalises minority-positive misclassifications more heavily to address class imbalance.
- Classifier: A linear-kernel SVM was selected for fast training on large-scale data and applied through LIBLINEAR.Preliminary kernel experiments supported a linear mapping.
- Model selection: Ten-fold cross-validation with grid search jointly evaluates feature-type combinations and hyperparameters, selecting the model by positive-class F1-score.The selected model is retrained on held-in data and evaluated on a randomly sampled 10% hold-out set.
- Baselines: Optimised models are compared with an unoptimised word-n-gram SVM and a keyword system based on aggressive-language and profanity lists.The keyword baseline marks a post positive when it contains a word from existing vocabulary lists.
- Features: The feature set combines word and character n-grams, term-list indicators, subjectivity features, and topic-model features.Character n-grams provide robustness to spelling variation in social-media data, while topic models represent categories such as threats, defamations, insults, and defenses.
5 Results
Experiments evaluated optimized linear-SVM systems for English and Dutch ASKfm posts using combined feature types, single feature types, and baseline comparisons. The systems outperformed keyword and word n-gram baselines, while errors remained concentrated in implicit, short, ironic, misspelled, and defensive posts.
- Overall performance: 64.26% English and 61.20% Dutch maximum cross-validation F1-scores were achieved by optimized systems combining varied feature types.The holdout results indicated good generalization, with little under- or overfitting.
- Overall performance: The optimized classifier clearly outperformed both keyword-matching and unoptimized word n-gram baselines.Keyword matching had the lowest performance in both languages, while feature-group and hyperparameter optimization substantially improved over the unoptimized word n-gram baseline.
- Feature analysis: Word n-grams, character n-grams, and subjectivity lexicons were strong individual or combined feature types for cyberbullying detection.Adding character n-grams consistently improved performance, whereas term lists performed poorly and topic models helped Dutch systems when combined with other features.
- Error analysis: Defense was the hardest category to classify, whereas threats, curses, and sexual harassment had the lowest error rates.Defensive posts from bystanders and victims often contained less aggressive language and were shorter than cyberbullying attacks.
- Error analysis: False negatives often lacked explicit cyberbullying vocabulary or were ironic, while short posts and posts with misspellings, grammatical errors, or incomplete words were also difficult to recognize.False positives often contained aggressive language directed at another person, including sexual or profane words.
6 Conclusions and Future Research
The paper investigates automatic detection of cyberbullying-related social media posts, extending coverage to bullies, victims, and bystanders. English and Dutch experiments support the feasibility of the approach while identifying finer-grained categories and participant roles as future directions.
- The research investigates automatic cyberbullying detection on social media to support rapid moderation when manual monitoring is infeasible.
- The system detects signals from posts written by bullies, victims, and bystanders, addressing roles often overlooked in prior attack-focused research.
- 64.32% and 58.72% F1-scores were achieved for English and Dutch classifiers, respectively, after feature selection and hyperparameter optimisation.
- The classifiers significantly outperformed keyword and unoptimised n-gram baselines on holdout tests of English and Dutch ASKfm data.
- Future work could detect fine-grained categories such as threats, curses, racism, and hate to prioritise severe cases for manual inspection.
- Future systems could identify participants’ roles to analyse incident context and severity and tailor moderation feedback to bullies, victims, or bystanders.