Social context in political stance detection: Impact and extrapolation

Area of Science:

• Computational Social Science
• Natural Language Processing
• Machine Learning

Background:

• Stance detection is crucial for analyzing public opinion and identifying harmful content like misinformation and hate speech.
• Existing models often struggle with generalizing to new contexts or future events.
• Social media platforms like Twitter offer rich data for understanding user stances and interactions.

Purpose of the Study:

• To evaluate the performance and extrapolation capabilities of political stance-detection models.
• To investigate the impact of incorporating user social networks (ego-networks) into stance detection.
• To assess model generalizability across different countries and time periods.

Main Methods:

• Utilized a large-scale, weakly-labeled Twitter dataset from the 2019 South American Protests.
• Developed transformer-based encoders for user and tweet embeddings.
• Employed heterogeneous graph attention networks (GATs) for stance prediction.
• Analyzed model performance in predicting user stances within and across different country contexts.

Main Results:

• Incorporating users' ego-networks consistently improved in-country political stance detection performance.
• Leveraging social context significantly enhanced the models' ability to extrapolate stance predictions to new country contexts.
• The models demonstrated improved generalizability to future events when social context was utilized.

Conclusions:

• Social network information is a valuable feature for enhancing the accuracy and generalizability of political stance detection models.
• Graph-based approaches, incorporating social context, offer a promising direction for robust stance detection.
• These findings have implications for improving opinion polling, misinformation detection, and understanding online political discourse.