Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Applying the Ego Network Model to Cross-Target Stance Detection

Jack Tacchi, Parisa Jamadi Khiabani, Arkaitz Zubiaga, Chiara Boldrini, Andrea Passarella

TL;DR

This work addresses Cross-Target Stance Detection under data-availability constraints by introducing two graph-based features from the Ego Network Model: ENM and SENM. Using node2vec embeddings, these features are integrated with RoBERTa and other social graph features in a few-shot, cross-target setting, showing macro F1 scores above 0.7 for most targets and shot levels. The study finds that outer (weaker but more numerous) ego circles are often more predictive than inner circles, and that signed vs unsigned ENMs yield similar performance, making ENM a robust data-efficient alternative to CT-TN. Overall, ENM/SENM offer practical CTSD solutions when CT-TN data are incomplete or inaccessible, supporting broader applicability of stance analyses in restricted data environments.

Abstract

Understanding human interactions and social structures is an incredibly important task, especially in such an interconnected world. One task that facilitates this is Stance Detection, which predicts the opinion or attitude of a text towards a target entity. Traditionally, this has often been done mainly via the use of text-based approaches, however, recent work has produced a model (CT-TN) that leverages information about a user's social network to help predict their stance, outperforming certain cross-target text-based approaches. Unfortunately, the data required for such graph-based approaches is not always available. This paper proposes two novel tools for Stance Detection: the Ego Network Model (ENM) and the Signed Ego Network Model (SENM). These models are founded in anthropological and psychological studies and have been used within the context of social network analysis and related tasks (e.g., link prediction). Stance Detection predictions obtained using these features achieve a level of accuracy similar to the graph-based features used by CT-TN while requiring less and more easily obtainable data. In addition to this, the performances of the inner and outer circles of the ENM, representing stronger and weaker social ties, respectively are compared. Surprisingly, the outer circles, which contain more numerous but less intimate connections, are more useful for predicting stance.

Applying the Ego Network Model to Cross-Target Stance Detection

TL;DR

This work addresses Cross-Target Stance Detection under data-availability constraints by introducing two graph-based features from the Ego Network Model: ENM and SENM. Using node2vec embeddings, these features are integrated with RoBERTa and other social graph features in a few-shot, cross-target setting, showing macro F1 scores above 0.7 for most targets and shot levels. The study finds that outer (weaker but more numerous) ego circles are often more predictive than inner circles, and that signed vs unsigned ENMs yield similar performance, making ENM a robust data-efficient alternative to CT-TN. Overall, ENM/SENM offer practical CTSD solutions when CT-TN data are incomplete or inaccessible, supporting broader applicability of stance analyses in restricted data environments.

Abstract

Understanding human interactions and social structures is an incredibly important task, especially in such an interconnected world. One task that facilitates this is Stance Detection, which predicts the opinion or attitude of a text towards a target entity. Traditionally, this has often been done mainly via the use of text-based approaches, however, recent work has produced a model (CT-TN) that leverages information about a user's social network to help predict their stance, outperforming certain cross-target text-based approaches. Unfortunately, the data required for such graph-based approaches is not always available. This paper proposes two novel tools for Stance Detection: the Ego Network Model (ENM) and the Signed Ego Network Model (SENM). These models are founded in anthropological and psychological studies and have been used within the context of social network analysis and related tasks (e.g., link prediction). Stance Detection predictions obtained using these features achieve a level of accuracy similar to the graph-based features used by CT-TN while requiring less and more easily obtainable data. In addition to this, the performances of the inner and outer circles of the ENM, representing stronger and weaker social ties, respectively are compared. Surprisingly, the outer circles, which contain more numerous but less intimate connections, are more useful for predicting stance.
Paper Structure (16 sections, 3 figures)

This paper contains 16 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Contributions
  3. Background
  4. Stance Detection
  5. The Ego Network Model
  6. Methodology
  7. Performing Cross-Target Stance Detection
  8. Feature Embeddings
  9. Model Hyperparameters
  10. Experimental Settings
  11. Computing Ego Networks
  12. Computing Ego Networks
  13. Generating Signed Connections
  14. Data
  15. Results
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Architecture of the CT-TN model.
  • Figure 2: The Ego Network Model, with the expected number of Alters of each circle (for humans).
  • Figure 3: Graphs displaying the performances of the individual text-based RoBERTa, the CT-TN model and the signed and unsigned Ego Networks features, as well as the unsigned inner and outer circle features. The shot number is displayed along the X axis and the averaged macro F1 scores along the Y axis.