Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

When Large Language Models Do Not Work: Online Incivility Prediction through Graph Neural Networks

Zihan Chen, Lanyu Yu

TL;DR

The paper tackles online incivility detection by moving beyond text-only models to a graph-based framework that jointly leverages linguistic content and relational context. It constructs a semantic-similarity graph over comments using Sentence-BERT embeddings and couples a GNN with an MLP through a dynamic attention-based fusion to adaptively weight nodal versus topological information. Across three incivility dimensions (personal attacks, aggression, toxicity) on the Wikipedia Detox dataset, the approach consistently outperforms twelve state-of-the-art LLM baselines while achieving lower inference costs, with the strongest gains in context-dependent categories. These findings highlight the value of incorporating structural context for scalable, efficient moderation in online communities and point to future work in enriching graph signals and expanding datasets.

Abstract

Online incivility has emerged as a widespread and persistent problem in digital communities, imposing substantial social and psychological burdens on users. Although many platforms attempt to curb incivility through moderation and automated detection, the performance of existing approaches often remains limited in both accuracy and efficiency. To address this challenge, we propose a Graph Neural Network (GNN) framework for detecting three types of uncivil behavior (i.e., toxicity, aggression, and personal attacks) within the English Wikipedia community. Our model represents each user comment as a node, with textual similarity between comments defining the edges, allowing the network to jointly learn from both linguistic content and relational structures among comments. We also introduce a dynamically adjusted attention mechanism that adaptively balances nodal and topological features during information aggregation. Empirical evaluations demonstrate that our proposed architecture outperforms 12 state-of-the-art Large Language Models (LLMs) across multiple metrics while requiring significantly lower inference cost. These findings highlight the crucial role of structural context in detecting online incivility and address the limitations of text-only LLM paradigms in behavioral prediction. All datasets and comparative outputs will be publicly available in our repository to support further research and reproducibility.

When Large Language Models Do Not Work: Online Incivility Prediction through Graph Neural Networks

TL;DR

The paper tackles online incivility detection by moving beyond text-only models to a graph-based framework that jointly leverages linguistic content and relational context. It constructs a semantic-similarity graph over comments using Sentence-BERT embeddings and couples a GNN with an MLP through a dynamic attention-based fusion to adaptively weight nodal versus topological information. Across three incivility dimensions (personal attacks, aggression, toxicity) on the Wikipedia Detox dataset, the approach consistently outperforms twelve state-of-the-art LLM baselines while achieving lower inference costs, with the strongest gains in context-dependent categories. These findings highlight the value of incorporating structural context for scalable, efficient moderation in online communities and point to future work in enriching graph signals and expanding datasets.

Abstract

Online incivility has emerged as a widespread and persistent problem in digital communities, imposing substantial social and psychological burdens on users. Although many platforms attempt to curb incivility through moderation and automated detection, the performance of existing approaches often remains limited in both accuracy and efficiency. To address this challenge, we propose a Graph Neural Network (GNN) framework for detecting three types of uncivil behavior (i.e., toxicity, aggression, and personal attacks) within the English Wikipedia community. Our model represents each user comment as a node, with textual similarity between comments defining the edges, allowing the network to jointly learn from both linguistic content and relational structures among comments. We also introduce a dynamically adjusted attention mechanism that adaptively balances nodal and topological features during information aggregation. Empirical evaluations demonstrate that our proposed architecture outperforms 12 state-of-the-art Large Language Models (LLMs) across multiple metrics while requiring significantly lower inference cost. These findings highlight the crucial role of structural context in detecting online incivility and address the limitations of text-only LLM paradigms in behavioral prediction. All datasets and comparative outputs will be publicly available in our repository to support further research and reproducibility.

Paper Structure

This paper contains 18 sections, 10 equations, 1 figure, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Method
  4. Graph Construction via Semantic Similarity
  5. Hybrid Graph Neural Network Architecture
  6. Graph Neural Network Branch
  7. Multi-Layer Perceptron Branch
  8. Dynamic Attention-Based Fusion
  9. Classification and Loss Function
  10. Experiment
  11. Data
  12. Model Configuration and Evaluation Metrics
  13. Results
  14. Personal Attack Detection:
  15. Aggression Detection:
  16. ...and 3 more sections

Figures (1)

  • Figure 1: The high-level architecture of the proposed framework. Semantic embeddings are first generated via Sentence-BERT to construct a similarity graph. The model utilizes dual branches (GNN and MLP) fused via dynamic attention to classify comments into incivility categories.