Entity alignment for multilingual knowledge fusion suffers from insufficient granularity in feature modeling and limited exploitation of structural information. An entity alignment method was proposed that integrated multi-level embedding features with a sparse matrix propagation mechanism. Entities were represented through a unified embedding that fused character-level features, word-level embeddings, and neighborhood relational information, enabling fine-grained semantic and structural expression. To promote efficient knowledge propagation, a sparse adjacency matrix was constructed based on relation embeddings, and a normalization-based mechanism was introduced to stabilize feature transmission across graphs. To enhance global consistency during alignment, Sinkhorn regularization was applied to refine the similarity matrix, followed by the Hungarian algorithm to obtain optimal one-to-one matching. Stable performance was achieved on multiple cross-lingual knowledge graph datasets in terms of evaluation metrics such as hit rate and mean reciprocal rank. Compared with representative methods such as SNGA and EAMI, the proposed approach demonstrated strong competitiveness, validating its accuracy and robustness.