A System to Monitor Cyberbullying based on Message Classification and Social Network Analysis

Stefano Menini, Giovanni Moretti, Michele Corazza, Elena Cabrio, Sara Tonelli, Serena VillataACL 2019

TL;DR

This paper introduces a system that combines social network analysis with message classification to detect cyberbullying on platforms like Instagram. It addresses the challenge of identifying repeated, targeted harassment by first reconstructing local communities and then flagging offensive messages within those communities.

What Problem Does This Paper Address?

Cyberbullying is a significant problem on social media, particularly for young users. While Natural Language Processing (NLP) techniques can detect abusive language, identifying cyberbullying at scale is difficult because it often involves repeated attacks against a specific victim. Existing systems often operate as "black boxes," which can be problematic. This paper aims to create a more transparent system that helps educators and supervisors monitor potential cyberbullying by visualizing both the social network structure and the content of messages exchanged within identified communities.

Methods

The system has two main components:

  • Network Extraction: This involves reconstructing local communities of high school students on Instagram. Since Instagram doesn't have official school pages, the authors exploit geo-tagging data from photos. They start with users who have posted from specific school locations (first layer) and then expand to users who interact with them through comments (second and third layers). This process builds an inclusive network, which is then "pruned" to isolate smaller communities of students using a modularity function and by analyzing user biographies for relevant cues (e.g., age, school year).
  • Message Classification: A modular neural architecture, specifically a Long Short-Term Memory (LSTM) network, is used for binary classification of messages as either containing abusive language or not. The model is trained using English FastText embeddings and hyperparameters that have shown good performance in previous hate speech detection tasks.

Datasets

  • Training Data: The paper mentions using a dataset described in Waseem and Hovy (2016), containing 16,000 English tweets manually annotated for hate speech. They also refer to datasets used in shared tasks for Italian and German hate speech detection.
  • Evaluation Data: A test set of 900 Instagram comments was manually annotated by the authors, comprising 787 non-offensive and 113 offensive messages.

Key Insights and Ideas

  • Geo-tagging as a proxy for community: Leveraging geo-tagged posts is a clever way to infer school-related communities on platforms like Instagram, which lack direct institutional affiliations.
  • Layered network expansion: Building the network by progressively adding users who comment on existing users' posts helps capture relevant social connections beyond direct followers.
  • Pruning for focus: The iterative pruning of the network is crucial for isolating the specific student communities where cyberbullying is most likely to occur, making the analysis more manageable and relevant.
  • Transparency in monitoring: The system's interface, which visualizes both network structure and flagged messages, aims to provide a transparent tool for educators, allowing them to understand the context of potential cyberbullying.

Key Results

  • The message classification module achieved a micro-averaged F1 score of 0.823 on the test set.
  • After pruning, the network was reduced to a core of 892 users and 2,435 edges, representing 14,565 messages.
  • The system's interface allows for two views: one for network exploration and another that highlights offensive messages in red.

Limitations and Future Work

  • Data limitations: The system relies on public user profiles, meaning private content is inaccessible. This limits the detection of cyberbullying to less than 1% of messages in their use case, whereas other studies report higher rates of offensive language.
  • Real-time processing: The current workflow for extracting and processing data is not real-time, with an estimated processing time of around one week for large networks.
  • Cross-lingual capabilities: While the system supports English and Italian, future work includes improving the classifier by extending the training set with more Instagram data and experimenting with cross-lingual strategies.
  • Broader application: The authors suggest the approach can be applied to other social networks and chat applications.

Summary updated

BibTeX
@inproceedings{menini2019system,
  author = {Stefano Menini and Giovanni Moretti and Michele Corazza and Elena Cabrio and Sara Tonelli and Serena Villata},
  title = {{A System to Monitor Cyberbullying based on Message Classification and Social Network Analysis}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}

Loading related content...

Figure