This paper addresses the problem of entity alignment (EA), which identifies identical real-world entities between knowledge graphs. Existing EA models utilize a pseudo-labeling strategy that repeatedly adds highly reliable predicted unaligned entity pairs to the initial alignment data to address the lack of initial alignment data for learning. However, this process overlooks the negative impact of confirmation bias, which leads to performance degradation. In this paper, we propose UPL-EA, an integrated pseudo-labeling framework that systematically addresses confirmation bias in pseudo-labeling-based EA by explicitly removing pseudo-labeling errors and improving entity alignment accuracy. UPL-EA consists of two components: (1) Optimal Transport (OT)-based pseudo-labeling reduces false matching between two knowledge graphs and effectively determines entity correspondences using discrete OT modeling. It also presents an effective criterion for inferring pseudo-labeling alignments that satisfy one-to-one correspondences. (2) Parallel pseudo-labeling ensemble improves pseudo-labeling alignment by combining predictions from multiple models trained independently in parallel. The improved pseudo-labeling alignment is then added to the initial alignment data to enhance model learning for subsequent alignment inference. The effectiveness of UPL-EA in removing pseudo-labeling errors has been verified theoretically and experimentally. Through extensive experimental results and in-depth analysis, we demonstrate that UPL-EA outperforms 15 competing baseline models and is useful as a general pseudo-labeling framework.
