Recently, contrastive learning has shown great potential in text-based Knowledge Graph Completion (KGC). However, two challenges remain: (1) the false negative problem, where the sampled negative training entity set may contain correct entities; and (2) how to efficiently generate high-quality hard negatives in contrastive learning. To address these issues, we propose a text-based method, KG-MPH, which introduces an embedding-based method to identify potential positive examples within the negative set, alleviating the false negative problem. Additionally, it leverages known positive examples to efficiently generate hard negatives. Finally, inspired by our solution to the first problem, we explore the use of potential positive examples to improve existing model evaluation methods, such as enhancing the matching scores for entities that are more likely to be correct answers. Extensive experiments demonstrate that KG-MPH further improves the performance of text-based methods on the WN18RR and FB15K-237 datasets.