Introduction
Measuring abstract sociological concepts like "soft power"—the ability of a social entity to influence others through appeal and attraction rather than coercion—has long been a challenge for researchers. Traditional social science relies on human experts to code texts, a process that is contextually rich but lacks scalability and objectivity. Conversely, modern computational social science uses machine learning to process big data, offering scale but often lacking the nuanced conceptual grounding required for valid sociological measurement.
This paper proposes a hybrid "intelligence fusion" framework. By combining top-down domain expertise with bottom-up word-embedding geometric projections, the authors create a scalable, reliable, and valid method for quantifying the soft power of entities such as entrepreneurs and global cities.
The Proposed Framework
The core of the research is a five-layer semantic mining framework designed to bridge the gap between human conceptualization and data-driven operationalization.
The framework operates through three primary stages:
- Conceptualization (Top-Down): Domain experts define the dimensions of soft power (e.g., "innovation," "leadership") and identify keywords representing both these dimensions and the social entities being measured.
- Weight Calculation: The system processes massive media datasets to calculate the importance of entity keywords. It utilizes a metric called TF*PDF (Term Frequency * Proportional Document Frequency), which can be augmented with engagement data like "thumbs-ups" to better reflect actual public impact.
- Geometric Projection (Bottom-Up): Using word embeddings (like Word2Vec), words are mapped into a high-dimensional semantic space. The soft power score is derived from the cosine similarity between the entity's vector and the dimension's vector. A ReLU activation function is applied to ensure that only positive semantic relationships (angles < 90 degrees) contribute to the final score.
Case Studies and Validation
The authors validated the framework using two distinct datasets representing different scales and data types.
Case 1: Chinese Elite Entrepreneurs
Using a dataset of 196 viral WeChat articles (including nearly 500,000 "thumbs-ups"), the framework measured the "national image carrying capacity" of top Chinese entrepreneurs. The model successfully ranked Jack Ma as the leading figure (score: 0.5396), demonstrating the framework's ability to process full-text data and engagement metrics efficiently—the entire pipeline ran in under three minutes on a standard CPU.
Case 2: Global City Branding
The framework was also applied to the GDELT Global News Knowledge Graph, analyzing ~11.5 million metadata records for 10 Alpha+ global cities. Despite using metadata and pre-trained embeddings rather than raw text, the model identified New York (1.69) and London (1.35) as the dominant global brands. These results perfectly align with established GaWC tier rankings, validating the model's accuracy.
Key Insights
- Semantic Geometry: Spatial relationships in neural networks (cosine similarity) can serve as a literal "projection" of sociological influence, translating abstract concepts into measurable indices.
- The Power Law of Influence: The research empirically confirmed a "long-tail" distribution of soft power, where a "vital few" entities carry the vast majority of a group's or nation's image.
- Adaptability: The framework is robust enough to function with either raw text or metadata/knowledge graphs, making it a versatile tool for researchers with varying data access.
Conclusion
By integrating human expertise with the computational power of word embeddings, this framework offers a rigorous yet scalable solution for measuring abstract social concepts. While the authors note that pre-trained models may sacrifice some specificity compared to custom-trained ones, the framework's high validity across different domains marks a significant step forward in computational social science.
Summary updated