Abstract
Reference linking is a fundamental component of bibliometrics, yet current state-of-the-art methods often fail to link nearly 50% of references due to rigid title-matching and dependency on specific target databases. This paper introduces a novel approach that treats reference linking as an entity resolution problem using "blocking" techniques. By implementing a three-step pipeline—pre-processing, token blocking with meta-blocking, and granular matching—the proposed system achieves a 90% increase in papers linked to target collections and a nine-fold increase in in-text citation coverage, maintaining a high F1-score of 85.71%.
Introduction
Reference linking—the process of connecting a citation in a paper to the actual publication it refers to—is the backbone of scholarly communication. It enables the calculation of h-indexes, the detection of research trends, and the construction of recommendation engines. However, existing methods are often "leaky," failing to resolve a significant portion of citations. The authors identify two primary culprits:
- Target-Collection Dependency: Most methods only link a reference if the cited paper already exists in a specific database (e.g., DBLP or OpenAlex).
- Brittle Matching: Systems relying on exact string matching fail when citations omit titles (common in Physics) or use non-standard abbreviations.
Methods
To overcome these limitations, the authors pivot toward Entity Resolution (ER) techniques, specifically Blocking. Comparing every reference to every other reference is computationally prohibitive ($O(N^2)$ complexity). To ensure scalability, the authors developed a three-step pipeline:
- Pre-processing: Raw citation strings are parsed using GROBID, and journal/conference abbreviations are normalized using a mapping of over 47,000 titles.
- Blocking: The core innovation involves Token Blocking to group references sharing common words. To maintain efficiency, Meta-blocking is applied to build a graph of entities and prune edges unlikely to represent matches.
- Matching: Within refined blocks, the system performs granular comparisons using a weighted average of Jaccard similarities across seven fields: title, author, journal, booktitle, year, volume, and pages.
Key Insights and Ideas
- Reference-to-Reference Linking: By matching references to each other rather than just to a target database, the system establishes "bibliographic coupling." This allows researchers to know that two papers cite the same work, even if that work is not yet indexed in a major database.
- Schema-Agnostic Robustness: Token blocking prevents the system from being "blinded" by missing fields. Matches can still be identified through combinations of authors, volumes, and page numbers.
- Optimized Meta-Blocking: The authors found that the Enhanced Common Blocks Scheme (ECBS) weighting, combined with Cardinality Node Pruning (CNP), provided the optimal balance of speed and accuracy.
Key Results
The study utilized unarXive, evaluating 300,000 recent references from nearly 10,000 papers. The results demonstrate massive improvements over baseline metadata:
- 90% Increase in papers successfully linked to the target collection.
- 5.6x Increase in bibliographic coupling (identifying papers citing the same sources).
- 9.6x Increase in the coverage of in-text citation markers.
- High Quality: Manual evaluation of 500 random pairs yielded a Precision of 93.2% and a Recall of 79.3%.
Limitations and Future Work
Despite these gains, challenges remain. The evaluation was primarily focused on STEM fields (Physics, Math, and Computer Science); the humanities, with their varied citation styles, represent a future challenge. Additionally, while blocking improves scalability, the authors plan to explore Progressive Blocking to prioritize the most likely matches in time-constrained environments. Future applications include using these enhanced links to improve full-text analysis and scholarly recommendation systems.
Summary updated