Entity alignment, a critical process for integrating knowledge across diverse knowledge graphs, involves identifying and matching entities with identical semantics. The conventional approach relies on semi-supervised learning to align entities based on the similarity of their embeddings. However, the heterogeneity of data sources often results in non-isomorphic neighborhood structures for aligned entities, posing significant challenges, particularly for rare and sparsely connected entities. This paper introduces a novel soft label propagation framework that leverages multi-source data and iterative seed enhancement. The framework utilizes seed entities as anchors and selects optimal relational pairs to generate soft labels enriched with neighborhood features and semantic relationship information. Furthermore, considering the heterogeneity in neighborhood structures, a novel bidirectional weighted loss function is introduced to reduce the distance between positive pairs and to manage antagonistic pairs distinctively. This approach outperforms current semi-supervised methods, enhancing entity alignment accuracy across various datasets.