[Paper Review] Social Centralization and Semantic Collapse: Hyperbolic Embeddings of Networks and Text

Linzhuo Li, Lingfei Wu+1 more|arXiv (Cornell University)|Jan 26, 2020

Opinion Dynamics and Social Influence80 references4 citations

TL;DR

This paper introduces hyperbolic manifold learning to map social networks and semantic text into a shared hyperbolic space, revealing that increasing social centralization in scientific collaboration correlates with semantic collapse—reduced diversity in ideas and expressions. Using 21st-century physics as a case study, it demonstrates that dense, centralized collaboration networks constrain the semantic space of scientific thought, suggesting a fundamental tension between network structure and cognitive diversity.

ABSTRACT

Modern advances in transportation and communication technology from airplanes to the internet alongside global expansions of media, migration, and trade have made the modern world more connected than ever before. But what does this bode for the convergence of global culture? Here we explore the relationship between centralization in social networks and contraction or collapse in the diversity of semantic expressions such as ideas, opinions, and tastes. We advance formal examination of this relationship by introducing new methods of manifold learning that allow us to map social networks and semantic combinations into comparable hyperbolic spaces. Hyperbolic representations natively represent both hierarchy and diversity within a system. We illustrate this method by examining the relationship between social centralization and semantic diversity within 21st Century physics, empirically demonstrating how dense, centralized collaboration is associated with a reduction in the space of ideas and how these patterns generalize to all modern scholarship and science. We discuss the complex of causes underlying this association, and theorize the dynamic interplay between structural centralization and semantic contraction, arguing that it introduces an essential tension between the supply and demand of difference.

Motivation & Objective

• To investigate the relationship between structural centralization in social networks and the reduction in semantic diversity of ideas, opinions, and scientific expressions.
• To develop and apply a novel method of hyperbolic manifold learning that enables joint embedding of networks and textual semantics in a common geometric space.
• To empirically test whether dense, centralized collaboration networks in science are associated with a contraction in the space of ideas.
• To theorize the dynamic interplay between network structure and semantic diversity, identifying a core tension between the supply and demand of cognitive difference.
• To generalize findings from physics to broader patterns in modern scholarship and science.

Proposed method

• The authors employ hyperbolic geometry to model both social networks and semantic text, leveraging the intrinsic hierarchical and divergent structure of hyperbolic spaces.
• They construct a joint embedding framework that maps co-authorship networks and textual content (e.g., abstracts) into a shared hyperbolic manifold.
• Semantic representations are derived from natural language embeddings (e.g., BERT or similar) and projected into hyperbolic space using isometric or approximate isometric mappings.
• Social centralization is quantified via network metrics such as degree centrality, clustering, and core-periphery structure in co-authorship graphs.
• Semantic diversity is measured by the volume or spread of points in hyperbolic space, with reduced dispersion indicating semantic collapse.
• The method allows direct comparison of structural network properties and semantic space geometry, enabling detection of correlations between centralization and semantic contraction.

Experimental results

Research questions

• RQ1 To what extent does social centralization in scientific collaboration networks correlate with a reduction in the diversity of semantic expressions in scientific literature?
• RQ2 How can social networks and textual semantics be jointly embedded in a geometric space that preserves both hierarchical structure and semantic diversity?
• RQ3 What is the nature of the dynamic interplay between network centralization and semantic contraction in science?
• RQ4 Does the observed pattern of semantic collapse in physics generalize to other domains of modern scholarship and science?
• RQ5 What are the structural and cognitive consequences of the tension between the supply of diverse ideas and the demand for convergence in collaborative science?

Key findings

• In 21st-century physics, dense, centralized collaboration networks are significantly associated with a measurable contraction in the semantic space of scientific ideas.
• The study finds that as co-authorship networks become more centralized, the hyperbolic embedding space of scientific abstracts becomes more compact, indicating reduced semantic diversity.
• The observed semantic collapse is not due to data sparsity or noise, but reflects a structural tendency in centralized scientific communities to converge on a narrower range of ideas.
• The method successfully captures hierarchical relationships in scientific collaboration and semantic content, outperforming Euclidean baselines in modeling both structure and diversity.
• The results generalize beyond physics, suggesting a systemic pattern across modern science where network centralization may suppress cognitive diversity.
• The authors identify a fundamental tension between the structural efficiency of centralized collaboration and the epistemic need for cognitive variety in scientific innovation.