Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

STEntConv: Predicting Disagreement with Stance Detection and a Signed Graph Convolutional Network

Isabelle Lorge, Li Zhang, Xiaowen Dong, Janet B. Pierrehumbert

TL;DR

STEntConv addresses the challenge of predicting disagreement between social media post authors by constructing a signed user–entity graph that captures stance toward entities and applying a weighted Signed Graph Convolutional Network. Stance toward entities is extracted in an unsupervised fashion using Sentence-BERT, enabling domain-agnostic graph construction without user history. On the Reddit DEBAGREEMENT dataset, STEntConv yields notable gains over text-only baselines, particularly on the (comment, reply) subset where target entities appear, and demonstrates robustness to ablations. The approach provides interpretable signals by linking disagreement to entity-level stances and can be extended to include common nouns for broader coverage across topics.

Abstract

The rise of social media platforms has led to an increase in polarised online discussions, especially on political and socio-cultural topics such as elections and climate change. We propose a simple and novel unsupervised method to predict whether the authors of two posts agree or disagree, leveraging user stances about named entities obtained from their posts. We present STEntConv, a model which builds a graph of users and named entities weighted by stance and trains a Signed Graph Convolutional Network (SGCN) to detect disagreement between comment and reply posts. We run experiments and ablation studies and show that including this information improves disagreement detection performance on a dataset of Reddit posts for a range of controversial subreddit topics, without the need for platform-specific features or user history.

STEntConv: Predicting Disagreement with Stance Detection and a Signed Graph Convolutional Network

TL;DR

STEntConv addresses the challenge of predicting disagreement between social media post authors by constructing a signed user–entity graph that captures stance toward entities and applying a weighted Signed Graph Convolutional Network. Stance toward entities is extracted in an unsupervised fashion using Sentence-BERT, enabling domain-agnostic graph construction without user history. On the Reddit DEBAGREEMENT dataset, STEntConv yields notable gains over text-only baselines, particularly on the (comment, reply) subset where target entities appear, and demonstrates robustness to ablations. The approach provides interpretable signals by linking disagreement to entity-level stances and can be extended to include common nouns for broader coverage across topics.

Abstract

The rise of social media platforms has led to an increase in polarised online discussions, especially on political and socio-cultural topics such as elections and climate change. We propose a simple and novel unsupervised method to predict whether the authors of two posts agree or disagree, leveraging user stances about named entities obtained from their posts. We present STEntConv, a model which builds a graph of users and named entities weighted by stance and trains a Signed Graph Convolutional Network (SGCN) to detect disagreement between comment and reply posts. We run experiments and ablation studies and show that including this information improves disagreement detection performance on a dataset of Reddit posts for a range of controversial subreddit topics, without the need for platform-specific features or user history.
Paper Structure (21 sections, 5 equations, 5 figures, 5 tables)

This paper contains 21 sections, 5 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Stance
  4. Signed Graphs
  5. Dataset
  6. Framework
  7. User-Entity Graph Construction
  8. STEntConv
  9. Baselines
  10. BERT
  11. GCN only
  12. StanceRel
  13. FALCON
  14. Training
  15. Results
  16. ...and 6 more sections

Figures (5)

  • Figure 1: User-entity graph visualised with Gephi bastian2009gephi (positive edges). We apply a force atlas layout. Pink nodes are entities, blue nodes are users which we can see clustered around the target entities they expressed a positive stance for.
  • Figure 2: Example user-entity graph. The network is signed, with each edge representing user stance towards an entity.
  • Figure 3: Model architecture.
  • Figure 4: Confusion matrix (STEntConv)
  • Figure 5: heatmap of cosine similarities between each subreddit name and target named entities.