Practical Author Name Disambiguation under Metadata Constraints: A Contrastive Learning Approach for Astronomy Literature

TL;DR

This work presents NAND, a neural, contrastive-learning approach for author name disambiguation that operates with minimal metadata (author name, title, abstract) to scale across large digital libraries. It introduces LSPO, a large NASA/ADS–ORCiD linked dataset designed to reflect real-world ambiguity in physics and astronomy, and demonstrates near 95% F1 clustering and over 94% pairwise accuracy. The method leverages Char2Vec or Word2Vec for name representations and transformer-based document embeddings (Specter, Llama 3.2) within a Siamese architecture, trained with InfoNCE or cosine embedding loss, followed by DBSCAN clustering within name blocks. Across in-domain LSPO and cross-domain LAGOS-AND benchmarks, NAND shows strong performance under sparse metadata and provides a practical, scalable solution for disambiguating author identities in large digital libraries.

Abstract

The ability to distinctly and properly collate an individual researcher's publications is crucial for ensuring appropriate recognition, guiding the allocation of research funding and informing hiring decisions. However, accurately grouping and linking a researcher's entire body of work with their individual identity is challenging because of widespread name ambiguity across the growing literature. Algorithmic author name disambiguation provides a scalable approach to disambiguating author identities, yet existing methods have limitations. Many modern author name disambiguation methods rely on comprehensive metadata features such as venue or affiliation. Despite advancements in digitally indexing publications, metadata is often unavailable or inconsistent in large digital libraries(e.g. NASA/ADS). We introduce the Neural Author Name Disambiguator, a method that disambiguates author identities in large digital libraries despite limited metadata availability. We formulate the disambiguation task as a similarity learning problem by employing a Siamese neural network to disambiguate author names across publications relying solely on widely available publication metadata-author names, titles and abstracts. We construct the Large-Scale Physics ORCiD Linked dataset to evaluate the Neural Author Name Disambiguator by cross-matching NASA/ADS publications ORCiD. By leveraging foundation models to embed metadata into features, our model achieves up to 94% accuracy in pairwise disambiguation and over 95% F1 in clustering publications into their researcher identities. We release the testing dataset as a benchmark for physics and astronomy, providing realistic evaluation conditions for future disambiguation methods. The Neural Author Name Disambiguator algorithm demonstrates effective disambiguation with minimal metadata, offering a scalable solution for name ambiguity in large digital libraries.

Figures (2)

• Figure 1: Top 150 name blocks with the highest number of unique ORCID identifiers in our dataset, represented by word size. One of the largest name blocks, Y. Wang, contains 153 unique ORCID identifiers across 426 author record instances. In contrast, the majority of the 97,936 name blocks (88,369) are associated with a single unique ORCID identifier.
• Figure 2: Overview of the pairwise similarity model training process and the clustering method: (1) Representing researcher information (e.g., author name) and publication content (e.g., title and abstract) using foundation models. (2) Training a Siamese neural network on the represented author records to distinguish if two author records refer to the same researcher or different researchers. (3) Clustering full name blocks of represented and transformed embeddings of author records utilizing DBSCAN with added constraints for naming conventions and years in-between publications.