A Flexible and Configurable System to Author Name Disambiguation

Abstract

Author Name Disambiguation (AND) is critical in maintaining the integrity of bibliographic databases, especially under data sparsity and large-scale ambiguity. This paper introduces a configurable and scalable AND system that combines transformer-based embeddings (MiniLM), Graph Convolutional Networks (GCN), and hierarchical clustering. The framework enables fine-grained parameterization of GCN depth, training epochs, and embedding models to adapt to datasets with varying structural and semantic complexity. Extensive evaluations on three benchmark datasets, including AMiner-12, DBLP, and LAGOS-AND, demonstrate consistent improvements over state-of-the-art baselines. On DBLP, our system achieves a pF1 of 0.878 and a K-Metric of 0.976, outperforming prior work by over 15% and 20%, respectively. On AMiner-12, despite sparse metadata, the method attains a 13.8% gain in average cluster purity and 10.1% in K-Metric. On the large-scale LAGOS-AND dataset, the system reaches a B-cubed F1-score of 0.908, surpassing the best-reported baseline by more than 9%. These results validate the system’s ability to integrate semantic and relational signals for robust and accurate AND across diverse contexts.