Entity alignment is the process of identifying and linking entities (like people, organizations, or locations) across different datasets or knowledge bases that refer to the same real-world object. Most entity alignment methods suffer from noise introduced when incorporating multi-hop neighborhood information and the neglect of interactions involving negative entity pairs, leading to suboptimal alignment performance. To address these limitations, we propose a Mixed Negative Sampling-based Entity Alignment model (MNS-EA). The model initially embeds entities into a unified vector space through multilingual training, subsequently employs a single-layer network to integrate one-hop neighborhood information, and constructs negative samples via a hybrid negative sampling algorithm to mitigate erroneous matches under information scarcity. Comparative experiments conducted on three DBP15K subsets (zh_en, ja_en, fr_en) demonstrate that the proposed model significantly enhances alignment accuracy. On the fr_en dataset, the proposed model demonstrates particularly notable metric improvements: Hits@1 achieves a 5.1% absolute gain over the suboptimal model HGCN, while Hits@10 outperforms RNMthe second-best baselineby 3.3%. Furthermore, comprehensive experiments conducted on the DWY100K dataset systematically validate both the component-wise effectiveness of our framework and its cross-dataset generalization capability.