Controversiality in Online Conversations

A multi-method analysis of ideological polarization and thread dynamics on DerStandard

Overview

Controversial discussions drive online discourse by attracting user engagement and shaping public opinion. This project presents a multi-method analysis of controversiality on DerStandard, an Austrian newspaper platform, integrating ideological positioning, network structure, and thread dynamics—an approach not previously combined in the literature.

Key Contributions

  • Multi-perspective integration: First study to combine ideological positioning, network structure, and thread dynamics for measuring online controversy
  • Large-scale German-language analysis: Analyzed a decade-long dataset (2013–2022) comprising 75 million user comments and 412 million up/down-votes
  • Controversiality metric: Developed a metric promoting balanced and sizeable inward positive and negative votes to assess user controversiality
  • Cross-method validation: Compared structural thread measures with vote-based metrics, revealing they capture distinct aspects of controversy

Methodology

The research is divided into two main analyses:

  1. Latent Ideology Extraction:
    • Identify influencers by ranking users with most positive, negative, and controversial incoming votes
    • Apply Correspondence Analysis (CA) to obtain user and influencer embeddings
    • Validate influencer classification using Large Language Models (LLMs)
    • Explore alternative approach using SHEEP embedding on user–user signed networks
  2. Thread-Level Controversiality:
    • Compare thread characteristics: depth, number of posts, negative votes
    • Compute controversy index based on h-index framework
    • Analyze author controversiality scores and inter-event response times
    • Measure polarization based on ideological stance of participating users

Results

The CA embeddings reveal a densely connected positive cluster aligned with isolated positive influencers, plus two larger clusters showing internal positive and cross-cluster negative interactions. Thread analysis shows strong alignment between structural measures of controversy, while vote-based metrics exhibit weaker correlations—suggesting these metrics capture distinct aspects of controversiality.

Technologies

  • Network Science
  • Correspondence Analysis
  • SHEEP Embedding
  • Large Language Models