Common Sense Reasoning for Detection, Prevention, and Mitigation of Cyberbullying (Extended Abstract)

Karthik Dinakar, Rosalind W. Picard, Henry LiebermanIJCAI 2015

Introduction

Traditional cyberbullying detection systems often fall short when faced with the nuances of human language. While explicit verbal abuse is relatively easy to flag using keyword filters, bullies frequently employ subtle, indirect methods that bypass standard profanity checks. This research explores a hybrid approach combining traditional machine learning with common sense reasoning to identify these "clean" but harmful insults.

The Challenge of Implicit Bullying

Most detection models rely on supervised learning to catch blatant harassment. However, human communication often relies on the "Maxim of Pragmatics"—the idea that we don't state the obvious. Bullies exploit this by using sarcasm or gendered stereotypes. For instance, telling a boy "you’ll be good at pressing my shirt" contains no slurs but uses a sexist stereotype to attack his masculinity. Without a baseline understanding of social norms, standard statistical models like SVMs or Naive Bayes often miss these context-heavy instances.

Methodology

The researchers developed a two-pronged detection strategy:

1. Supervised Learning for Explicit Abuse

Standard algorithms, including Naive Bayes, Rule-based JRip, Tree-based J48, and SVMs, were used to detect blatant harassment. The feature set included general linguistic markers and label-specific unigrams/bigrams.

Feature engineering for supervised learning

2. Common Sense Reasoning for Implicit Abuse

To address subtle bullying, the authors utilized ConceptNet, a semantic network of everyday knowledge. They built BullySpace, a specialized layer containing over 200 assertions about common stereotypes (e.g., "lipstick is used by girls"). Using AnalogySpace (based on Singular Value Decomposition), the system calculates the "cosine similarity" between comments and gendered concepts, identifying when a comment "cross-wires" expectations to create an insult.

Key Results

The study evaluated these models across datasets from YouTube and Formspring, focusing on categories like Sexuality, Race and Culture, and Intelligence.

Performance metrics for explicit bullying detection across different labels and models

As shown in the performance data, the rule-based JRip model achieved a notable 80% accuracy in detecting explicit bullying related to sexuality. While accuracy varied across labels, SVMs consistently showed high reliability, as indicated by their high kappa values (e.g., 0.79 for Sexuality).

Beyond statistical performance, the "BullySpace" model successfully flagged implicit insults that traditional models missed. In a human-AI agreement test involving 50 subtle instances, human annotators agreed with the common sense model's metrics in approximately 60-75% of cases.

Insights and Future Directions

A significant challenge remains "knowledge sparsity." If a cultural reference (e.g., a specific celebrity) is missing from the knowledge base, the reasoning chain fails. Future work aims to expand this database with contemporary references and move toward predictive modeling for adolescent distress. Additionally, the authors propose "reflective user interaction," such as introducing cooling-off delays when the system detects potential harassment, shifting the focus from mere detection to active mitigation.

Summary updated

BibTeX
@inproceedings{dinakar2015common,
  author = {Karthik Dinakar and Rosalind W. Picard and Henry Lieberman},
  title = {{Common Sense Reasoning for Detection, Prevention, and Mitigation of Cyberbullying (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure