BOND: Bootstrapping From-Scratch Name Disambiguation with Multi-task Promoting

TL;DR

BOND tackles the from-scratch name disambiguation problem by jointly learning local paper similarities and global clustering in an end-to-end framework. It constructs a multi-relational graph per name, uses a graph attention encoder–decoder to reconstruct local edges, and incorporates DBSCAN-derived pseudo-labels to steer cluster-aware learning, with these signals reciprocally reinforcing each other. The approach yields state-of-the-art results on WhoIsWho-v3, and its enhanced BOND+ variant with ensemble and post-match achieves top performance on the WhoIsWho leaderboard. The work demonstrates the value of end-to-end multi-task promoting for disambiguation tasks and provides insights into multi-relational graph design, clustering integration, and practical robustness considerations.

Abstract

From-scratch name disambiguation is an essential task for establishing a reliable foundation for academic platforms. It involves partitioning documents authored by identically named individuals into groups representing distinct real-life experts. Canonically, the process is divided into two decoupled tasks: locally estimating the pairwise similarities between documents followed by globally grouping these documents into appropriate clusters. However, such a decoupled approach often inhibits optimal information exchange between these intertwined tasks. Therefore, we present BOND, which bootstraps the local and global informative signals to promote each other in an end-to-end regime. Specifically, BOND harnesses local pairwise similarities to drive global clustering, subsequently generating pseudo-clustering labels. These global signals further refine local pairwise characterizations. The experimental results establish BOND's superiority, outperforming other advanced baselines by a substantial margin. Moreover, an enhanced version, BOND+, incorporating ensemble and post-match techniques, rivals the top methods in the WhoIsWho competition.

Figures (5)

• Figure 1: An overview of the SND problem and performance comparisons between BOND and baselines. (a) Paper connections are established through diverse relationships. Noise is observed in the linkage of Paper P4 to Paper P3; (b) SND-all*: Single Model Version of SND-all, pm.: post-match.
• Figure 2: The overall framework of BOND and other SND methods. NE: network embedding; HAC: hierarchical agglomerative clustering; our proposed framework, as depicted in (b), integrates metric learning and clustering within a multi-task learning framework. By optimizing the weighted sum of reconstruction loss and cluster-aware loss, the global information derived from the clustering component can reciprocally guide the local information extracted from the reconstruction part.
• Figure 3: Effect of different losses and multi-relational features.(a): All: training on all loss; All-recon: the clusters of local metric learning; All-cluster: the outputs of cluster-aware learning. (b): Training only on Local Metric Learning Loss. recon: the clusters of local metric learning; cluster: the outputs of cluster-aware learning. (c) and (d): A: CoAuthor; O: CoOrg; V: CoVenue.
• Figure 4: The results of ablation analysis of our ensemble model. (%) ">0" signifies that the threshold is set to 0. "CoA[0,1]" denotes the ensembling of models with coauthor values greater than 0 and 1, respectively. "CoA, CoO[0.5,0.6]" refers to the multi-view model with coauthor and coorg edges, when coauthor greater than 0 and 1 and coorg greater than 0.5 and 0.6, respectively.
• Figure 5: Analysis of hyper-parameters.

Theorems & Definitions (3)

• definition 1
• definition 2
• definition 3