PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Functional Embedding Models for Funder Name Disambiguation in Scientific Publication Records

Kanyao Han, Zhiwen You, Jinseok Kim, Jana Diesner

arXiv:2609.09984v1cs.CL

TL;DR

Large-scale funding analyses are hindered by inconsistent funder names and incomplete mappings to unique identifiers. This paper integrates publication and registry data to fine-tune multifunctional embedding models, achieving strong matching performance and extending disambiguation through clustering beyond registry coverage.

  • Problem

    Inconsistent spellings, translations, abbreviations, and incomplete mappings make it difficult to link funder-name occurrences to standardized unique identifiers for large-scale funding analysis.

  • Method

    The framework links Web of Science, Crossref Open Funder Registry, and Research Organization Registry data to create training pairs, then fine-tunes Sentence Transformer, Gemma, and Qwen3 embeddings with multi-task learning.

  • Results

    Fine-tuned embedding models substantially outperform pre-trained embedding models and generative LLMs, while the workflow also clusters funder names that are unlikely to be indexed in ROR.

  • Takeaways & Limitations

    The reusable framework supports multiple funder-disambiguation functions across model architectures and may apply beyond biodiversity conservation because its training data include funders from multiple research domains.

  • Takeaways & Limitations

    Lengthy prefixes or suffixes can cause correct matches to receive low similarity scores and be filtered out during a later classification step.

Abstract

from arXiv · show

Understanding the historical allocation and distribution of research funding advances our knowledge of how scientific research is supported across fields, institutions, and regions. However, large-scale analyses are hindered by the lack of comprehensive funder name disambiguation solutions, as funder names often exhibit spelling variations, translations, abbreviations, and inconsistent levels of granularity. In this paper, we present a framework for developing multilingual, multi-functional funder name disambiguation models and demonstrate its application to research publications in biodiversity conservation. To construct a training dataset, we integrated the Research Organization Registry (ROR), which provides unique identifiers for research organizations, with two publication datasets: the Web of Science (WoS) and the Crossref Open Funder Registry (OFR). We used multi-task learning with Contrastive Loss and Multiple Negatives Ranking Loss to fine-tune three open-weight embedding models from the Sentence Transformer, Gemma, and Qwen3 families. The best-performing models achieved accuracy above 0.90 when matching WoS funder names to ROR identifiers, outperforming general-purpose LLMs, including GPT-5.2, Claude-Sonnet-4.6, and Gemini-2.5-Flash, by more than 0.1. For funder names not indexed in ROR, we constructed a similarity network among funder names and identified clusters within it. Finally, we analyzed the disambiguation results and highlighted challenges arising from limited knowledge of smaller funders and funders from non-English-speaking countries. This work provides a reusable framework for funder name disambiguation with potential applicability across different model architectures and datasets, featuring cost-effective training data creation and multi-task learning and disambiguation.

1. Introduction

Large-scale funding-allocation analysis is limited by inconsistent funder names and incomplete standardized mappings. The paper addresses this gap with a reusable, multi-functional disambiguation framework built from linked publication and organizational data.

  • Motivation: Funder records vary through spelling differences, translations, abbreviations, and granularity, complicating links to unique identifiers.WoS funder names are often extracted from acknowledgments, where the same funder can appear under multiple names.
  • Method: The framework creates labeled name pairs by linking WoS and OFR records, then augments them with ROR name variants for model training.The workflow targets the limited availability of labeled funder-name pairs.
  • Results: Fine-tuned Sentence Transformer and Gemma models achieved 0.98 accuracy for ROR names and over 0.91 for WoS names, outperforming general-purpose LLMs by more than 0.1.The models use Contrastive Loss and Multiple Negatives Ranking Loss.
  • Results: 37.4% of WoS funder names were not indexed in ROR, so the study constructed similarity networks and applied Louvain clustering to identify funder groups beyond ROR.The analysis highlights challenges involving programs and funders from non-English-speaking countries.
  • Contributions: The contribution is a reusable framework spanning training-data creation, fine-tuning, evaluation, and multiple disambiguation tasks across model architectures.Supported tasks include same-funder classification, candidate matching, and clustering.

2. Related Work

Prior disambiguation work relies on matching, clustering, shared components, rules, and curated resources, but funder names present broader variation than institution names. Task-specific fine-tuning therefore remains an underexplored direction.

  • Existing approaches: Institution-name disambiguation commonly matches records to identifier dictionaries or groups records into clusters representing entities.Both approaches require deciding whether similar names refer to the same entity.
  • Existing approaches: Earlier methods use shared words, characters, locations, email components, edit distances, author information, and handcrafted rules.These methods combine lexical and contextual evidence to group institution names.
  • Funder disambiguation: A prior large-scale pipeline linked WoS funder names to OpenAlex and ROR using normalization, clustering, rules, named-entity recognition, and manual validation.It assigned at least one potential match to 1.9 million of 7.4 million unique funder names, covering 72% of funder mentions.
  • Research gap: Institution-name rules may not generalize to funders because institution records contain structured contextual components, while funder names have more diverse resolution and organizational forms.Funder records can include governmental departments, nongovernmental organizations, businesses, and funding programs.
  • Research gap: The paper positions task-specific fine-tuning of pretrained models as an underexplored approach for funder name disambiguation.The application focuses on biodiversity conservation publications.

3. Data

The study combines complementary datasets to improve funder information for biodiversity-conservation analyses. WoS offers broad but variable funder coverage, while OFR and ROR provide more standardized records and identifiers.

  • WoS contains 122,508 biodiversity-related conservation publications downloaded as of February 3, 2022, with funding data recorded since 2008.Funding-related data from 2008 onward are generally more complete than comparable Scopus and PubMed data.
  • WoS provides broader funder-record coverage than OFR, but its names contain more typos, variants, and inconsistent organizational resolution.OFR records are more standardized, yet omit a significant portion of funder records indexed in WoS.
  • The dataset combination creates a challenging benchmark because WoS heterogeneity and OFR standardization provide complementary coverage and naming information.The study uses WoS, OFR, and ROR to support funder-name disambiguation in biodiversity conservation.
  • ROR indexes more than 102,000 organizations, assigning each a unique ID, primary name, name variants, and organizational metadata.ROR also connects to other databases through funder identifiers such as Crossref Funder IDs.

4. Methodology

The methodology combines similarity-based mapping and clustering with a training set reconstructed from WoS, OFR, and ROR. Multi-task fine-tuning uses complementary losses to handle positive-negative imbalance and improve funder-name matching.

  • Workflow: The workflow comprises training-data creation, model fine-tuning, and application of the fine-tuned model for funder-name disambiguation.For ROR-indexed funders, names are mapped to registry records; other names are clustered by similarity.
  • Training Data Restructuring: Training pairs are created by linking WoS and OFR through paper DOIs, then using OFR as a bridge to connect names with ROR identifiers.Within-paper matching is used because WoS and OFR may list different funder names for the same paper.
  • Training Data Restructuring: Hard negatives are formed from each funder name’s 20 most similar names after known positive pairs are removed, while lower-similarity soft negatives are excluded.The procedure yielded 215,163 negative pairs from the WoS-derived data.
  • Training Data Restructuring: ROR expansion produced 140,684 positive and 3,061,200 negative pairs, with all positives and 500,000 randomly selected negatives retained for training.The ROR-derived pairs expand the range of names beyond the WoS-linked examples.
  • Multi-task Learning: Multi-task learning combines Contrastive Loss and Multiple Negatives Ranking Loss to address the approximately 3.7-to-1 negative-positive pair imbalance.Contrastive Loss separates dissimilar pairs using a margin, whereas Multiple Negatives Ranking Loss identifies one positive among K candidates.
  • Multi-task Learning: Contrastive Loss uses cosine distance between normalized embeddings, with margin m = 0.5 corresponding to an upper cosine-similarity threshold of 0.5.The binary label indicates whether a pair refers to the same funder entity or distinct entities.
  • Baselines: The evaluated baselines include English and multilingual Sentence Transformer models, lightweight EmbeddingGemma, and Qwen3-Embedding models.The multilingual Sentence Transformer supports more than 50 languages, while EmbeddingGemma has fewer than 0.5 billion parameters.

5. Matching Accuracy on Test Data

Matching accuracy varies with name cleanliness, funder frequency, model architecture, and whether models are fine-tuned. Fine-tuning substantially improves noisy WoS matching, while generative LLMs alone underperform embedding-based approaches.

  • Funder frequency: Accuracy is consistently higher on the complete WoS dataset than on its unique subset, suggesting frequent funders are easier to disambiguate.Common funders may contribute more fine-tuning examples and richer pre-trained representations.
  • Data cleanliness: Pre-trained embeddings achieve approximately 0.98 accuracy on standardized ROR names but only 0.76 for the best model on noisy WoS names.The results identify noisy, real-world names as the primary matching challenge.
  • Generative LLMs: Generative LLMs followed by exact matching achieve around 0.7 accuracy on ROR data and 0.6–0.67 on WoS data.They do not consistently produce canonical funder names required for reliable exact matching.
  • Hybrid approaches: Combining generative LLMs with pre-trained embeddings raises WoS accuracy from 0.76 to 0.81–0.82 but lowers standardized ROR performance relative to approximately 0.98.Incorrect standardization can propagate errors to downstream embedding matching.
  • Fine-tuned models: Fine-tuned ST and Gemma models reach approximately 0.91 accuracy on all WoS funder names and 0.89 on unique names.They outperform corresponding pre-trained models by 0.15–0.21 and 0.15–0.23, respectively.
  • Model size: Increasing embedding-model size does not consistently improve funder-name matching, as smaller fine-tuned ST and Gemma models outperform fine-tuned Qwen3-0.6B.Before fine-tuning, Gemma also outperforms larger Qwen3 models on WoS data.

6. Disambiguation on the WoS Corpus

The WoS corpus was disambiguated by matching funder names to ROR and clustering names that fell outside confident ROR coverage. Threshold evaluation quantified the matched subset, while clustering exposed long-tail, regional, and program-related coverage challenges.

  • Distinguishing Funder Names within ROR’s Coverage: A cosine similarity threshold of 0.85 produced the best balance, with precision 0.94, recall 0.90, and F1 0.92.The threshold classified names below 0.85 as unmatched and names at or above 0.85 as matched.
  • Distinguishing Funder Names within ROR’s Coverage: 93,155 funder names, or 62.6% of the WoS corpus, were matched to 7,953 unique ROR-indexed funders.The remaining 55,635 names represented 37.4% of the corpus.
  • Clustering Funder Names Beyond ROR’s Coverage: The remaining dataset was clustered using pairwise cosine similarity to support a second disambiguation round and diagnose why ROR matching failed.The clustering approach grouped unmatched names and helped identify funders absent from ROR.
  • Clustering Funder Names Beyond ROR’s Coverage: The remaining dataset was more right-skewed: 76.2% of its funders appeared once, compared with 46.8% in the main dataset.This distribution reflects many small, local, and specialized funders with limited representation in ROR.
  • Clustering Funder Names Beyond ROR’s Coverage: Seven of the ten most frequent funders in the clustering dataset were outside ROR’s scope, including major European and Chinese funding programs.ROR typically excludes programs that fall outside its organizational scope.
  • Clustering Funder Names Beyond ROR’s Coverage: Non-English-speaking funders created additional challenges because many were absent from ROR or recorded in languages other than English.Name variation and incomplete registry coverage complicated matching in this group.

7. Discussion and Conclusion

The paper contributes a reusable, multi-source framework for funder-name disambiguation while identifying important coverage and matching challenges. Its discussion emphasizes multi-task functionality, underrepresented funders, and cautious use of the resulting model and data.

  • Contributions: The framework combines publication records and organizational information to create training data for fine-tuning funder-name disambiguation models.It links sources such as WoS and OFR with ROR to support model training.
  • Contributions: Multi-task learning produces a multi-functional model for mapping names, classifying ROR matchability, and clustering names unlikely to be indexed in ROR.The model addresses several disambiguation steps because no single step resolves every challenge.
  • Challenges: Rare, small, local, specialized, and non-English-speaking-country funders remain difficult to disambiguate because they are underrepresented in publication records or curated resources.Differences in language, naming conventions, and funding systems can further limit representation in resources such as ROR.
  • Challenges: Clustering can help identify some frequently occurring funders that are absent from curated resources.This provides a partial approach for funders not represented in resources such as ROR.
  • Limitations and future work: Some correct matches may be filtered out when lengthy prefixes or suffixes lower similarity scores in a subsequent classification step.The paper suggests generative LLMs may assist with name cleaning or cluster verification rather than direct official-name prediction.
  • Cautious use: Users should treat the model and data cautiously because disambiguation errors remain possible and the broad ROR candidate pool may reduce matching accuracy.The paper recommends narrowing candidate organizations and removing acronyms before matching WoS names to ROR names.

8. Data and Code

The study’s reproducibility materials are constrained by the licensing terms governing its Web of Science-derived datasets.

  • Data and code: Clarivate’s licensing agreements prohibit redistributing raw publication records with the disambiguated funder names.The restriction applies to the Web of Science-derived datasets used in the study.
  • Data and code: The authors provide a mapping file containing publicly accessible article titles, WoS IDs, and DOIs for researchers with authorized WoS access.These identifiers allow authorized users to reconstruct the relevant records.

9. Funding Statement

The research was supported by the John D. and Catherine T. MacArthur Foundation through Grant No. 18-1802-152800-CSD.

  • Funding Statement: The John D. and Catherine T. MacArthur Foundation supported this research through Grant No. 18-1802-152800-CSD.
Loading 2609.09984v1…