Semantic and Structural Fusion for Academic Publishing Network Analysis Using OpenAlex Data

Overview

This work introduces a semantic approach to analyzing the academic publishing network, building upon a pre-existing structural methodology. The core innovation involves integrating a semantic layer, represented by citation-informed vector embeddings, with the established structural analysis. This integration facilitates a parameterized structural-semantic fusion designed to enhance understanding of relationships within the academic network.

Research Context

Prior work in analyzing the academic publishing network primarily focused on structural graph analysis, which captures topological relationships among entities without directly incorporating the content of academic works. This structural foundation, while effective for understanding network topology, inherently omits semantic insights derived from the textual content of documents. The current research aims to complement this structural view by adding a semantic dimension.

Approach

The research methodology involved several key steps to integrate semantic information with structural analysis:

• Document Vector Representations:

  Scientific documents were represented using citation-informed vector embeddings, specifically SPECTER2. These embeddings capture semantic information based on how documents are cited.

• Data Storage:

  The generated vector embeddings were stored in an embedded vector database. Each embedding was keyed by its stable OpenAlex ID, ensuring direct connectivity to the existing graph layer of the academic publishing network.

• Hybrid Structural-Semantic Fusion:

  A modular late-fusion function was defined to combine two types of similarity:
  • Semantic Similarity: Measured by the cosine of the SPECTER2 embeddings.
  • Structural Similarity: Measured using bibliographic coupling.
  This fusion function incorporated a tunable weight, alpha $(\alpha)$, allowing its value to be adjusted according to the specific task at hand.

• Corpus and Evaluation:

  The approach was evaluated on a corpus derived from VSB - Technical University of Ostrava. Two specific aspects were investigated:
  • The agreement of citation-informed embeddings with the expert OpenAlex topical taxonomy.
  • The performance of structural, semantic, and combined signals in a recommendation use case, assessing their information-carrying capacity under different data availability regimes.

• Open-Source Release:

  The entire approach has been released as an open-source extension to the existing apnet library, accompanied by a reproducible workflow.

Findings

The research yielded two primary findings from its application on the VSB - Technical University of Ostrava corpus:

• Embedding Alignment with Taxonomy:

  Citation-informed embeddings demonstrated better agreement with the expert OpenAlex topical taxonomy compared to a TF-IDF baseline. This suggests that SPECTER2 embeddings provide a more accurate semantic representation aligned with established academic classification systems.

• Task-Dependent Signal Utility:

  In a recommendation use case, the structural, semantic, and combined signals carried information in different regimes, depending on the available data. This indicates that while hybrid fusion is not universally superior, it provides an explicit mechanism for steering complementary signals to optimize for specific tasks. The value of combining these signals is contingent on the particular requirements and data context of the task.

Why This Matters

This work provides an open-source extension to the apnet library, offering a reproducible workflow for its semantic and structural fusion methodology. It establishes a direct connection between document vector representations and the graph layer via OpenAlex IDs, offering a concrete implementation for integrating content semantics into network analysis.