Reconnecting Fragmented Citation Networks with Semantic Augmentation

Abstract: Citation graphs are fundamental tools for modeling scientific structure, but are often fragmented due to missing citations of scientifically connected articles. To address this issue, we propose a computationally efficient hybrid framework integrating citation topology with LLM-based text similarity. Using 662,369 Web of Science publications in Mathematics and Operations Research & Management Science, we augment the original graph by adding semantic edges from small, disconnected components and weighting existing citations according to textual similarity. Semantic augmentation substantially reduces fragmentation while preserving disciplinary homogeneity. Compared to embedding-only clustering, cluster detection on augmented graphs using the Leiden algorithm retains structural interpretability while offering multi-scale organization. The method scales efficiently to large datasets and offers a practical strategy for strengthening citation-based indicators without collapsing disciplinary boundaries.

• The paper demonstrates that integrating LLM-based semantic embeddings with citation topology substantially reduces citation network fragmentation while preserving disciplinary integrity.
• The method employs k-nearest-neighbor repairs and cosine similarity weighting on citation edges, achieving robust results with reduced clusters and maintained homogeneity (>97% Mathematics, >85% OR&MS).
• The approach offers a scalable framework that improves bibliometric analysis by addressing missing links and decreasing computational overhead in large-scale citation datasets.

Semantic Augmentation for Fragment Repair in Citation Networks

Introduction

The study "Reconnecting Fragmented Citation Networks with Semantic Augmentation" (2605.12263) addresses methodological challenges inherent in relying solely on traditional citation graphs for mapping scientific domains. Citation networks, typically constructed from database records, often exhibit substantial fragmentation due to missing links induced by coverage limitations, time-window truncation, and extraction artifacts. These structural deficits undermine downstream bibliometric indicators, including community detection, interdisciplinarity measurement, and impact normalization. The paper proposes a hybrid framework integrating citation topology and semantic similarity obtained via LLM embeddings, demonstrating its efficacy in reducing fragmentation while retaining interpretability and disciplinary coherence.

Dataset and Baseline Structure

The authors utilize 662,369 publications from the Web of Science, spanning Mathematics and Operations Research & Management Science (OR&MS), after stringent filtering for metadata completeness and abstract length. The resulting citation network comprises 4,150,852 directed edges. Baseline clustering with the Leiden algorithm yields 33 clusters—dominated by two large communities corresponding to the disciplinary partitions but numerous small disconnected clusters. Importantly, cluster homogeneity is high: Mathematics at ~98% and OR&MS at ~86%. Analysis of reference retention reveals that only 17.1% of small cluster references are recovered in the extracted graph, establishing that fragmentation predominantly arises from incomplete connectivity rather than substantive intellectual boundaries.

Methods: Hybrid Semantic Augmentation

Semantic Augmentation

Semantic representations are generated by encoding concatenated titles and abstracts via the mxbai-embed-large LLM, yielding 1024-dimensional embedding vectors. The augmentation is implemented in two steps:

• Repair of Small Components (S1): k-nearest-neighbor (kNN) edges based on semantic similarity are introduced selectively for nodes in small clusters, with experiments over $k = 10, 100, 1000$.
• Semantic Weighting of Citation Edges (S2): Existing citation edges receive weights based on cosine similarity between textual embeddings, modulating the influence of citation links according to topical proximity.

Edges from S1 and S2 are merged using an adjustable parameter $a \in [0,1]$ for balancing the contribution from semantic and citation signals.

Community Detection and Comparison

Leiden clustering is applied to the augmented graphs, both weighted and unweighted, and compared to baseline and embedding-only clustering (K-means on embeddings). Evaluation criteria include number of clusters, cluster size distribution, disciplinary homogeneity, and stability.

Empirical Analysis and Strong Numerical Results

Fragmentation Reduction

Semantic augmentation demonstrates marked reduction in fragmentation:

• With $k = 10$, the cluster count drops from 33 to 9.
• With $k = 100$, cluster count reduces to 6.
• With $k = 1000$, only 2 clusters emerge, directly mapping to Mathematics and OR&MS.

Homogeneity remains robust, with Mathematics > 97% and OR&MS > 85%, across all augmentation levels. The complete absorption of small clusters into dominant disciplinary communities without significant loss of topical specificity is a strong numerical outcome.

Semantic Weighting Effects

Introducing semantic weights yields nuanced structural refinement. The weighted G1000nn graph produces four clusters, splitting out new small groups otherwise invisible in the unweighted graph. These reveal nodes with sparse reference connectivity and no semantic neighbors, suggesting further opportunities for augmentation.

Comparison With Embedding-Only Clustering

Embedding-only K-means clustering (fixed $k=2$) achieves comparable disciplinary homogeneity (Mathematics 98.3%, OR&MS 84.8%) but misclassifies a larger portion of Mathematics papers into OR&MS. It cannot discover multi-scale community structure or exploit citation topology. The hybrid approach adapts clustering granularity to both citation and semantic signals, preserving structural interpretability.

Computational Cost and Scalability

The hybrid augmentation drastically reduces computational overhead relative to fully dense semantic graphs. Repair mode restricts similarity computation to structurally vulnerable regions (small clusters), scaling at $O(S \log N)$ (S: small subset), while semantic weighting scales with the number of citation edges ($O(E)$). Overall, complexity is at most linear in node count and citation density, enabling application at Web of Science scale.

Implications and Future Directions

Theoretically, the work advances citation-based science mapping by demonstrating that selective semantic augmentation compensates for structural artifacts without eroding disciplinary partitions. Practically, the hybrid approach fortifies the robustness of bibliometric indicators, supporting more stable field delineations and impact normalization.

Further developments may include adaptive augmentation strategies for varying disciplinary architectures, enhanced weighting schemes integrating domain-specific LLMs, and application to other scientific areas where fragmentation distorts community detection. The methodology also provides a blueprint for scalable semantic-citation integration in algorithmic librarian frameworks and large-scale research evaluation platforms.

Conclusion

The paper establishes that citation networks, while epistemically curated, are structurally vulnerable to fragmentation artifacts. Targeted semantic augmentation, leveraging LLM embeddings, effectively repairs these gaps and preserves established disciplinary boundaries. Unlike embedding-only approaches, the hybrid framework retains topological interpretability and supports multi-scale analysis. These results provide a computationally feasible, methodologically transparent prescription for improving citation-based science mapping and bibliometric robustness at scale.