Social world knowledge: Modeling and applications.

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Network Science

Background:

• Factual world knowledge is widely represented in knowledge bases, but social world knowledge remains under-resourced.
• Effective communication and information processing require understanding social dynamics, which current knowledge bases do not capture.
• Existing entity embeddings focus on factual information, failing to represent social relationships.

Purpose of the Study:

• To develop a framework for eliciting low-dimensional entity embeddings from social contexts.
• To capture social aspects of world knowledge that are missed by fact-based embeddings.
• To create a valuable resource for understanding social world knowledge and its applications.

Main Methods:

• Developed SocialVec, a framework to learn entity embeddings from social network co-following data.
• Trained embeddings on approximately 200K entities using data from 1.3M Twitter users and their followed accounts.
• Evaluated embeddings on political bias assessment of news sources and personal trait prediction of users.

Main Results:

• SocialVec embeddings demonstrated advantageous or competitive performance on bias assessment and trait prediction tasks compared to baselines.
• Showcased that fact-based entity embeddings do not capture social knowledge effectively.
• The learned social entity embeddings successfully represent social relationships and user characteristics.

Conclusions:

• SocialVec provides a novel approach to represent social world knowledge through entity embeddings.
• The learned embeddings are valuable for tasks requiring understanding of social dynamics and relationships.
• The released embeddings will facilitate further research in social world knowledge and its applications.