Knowledge Graph Embedding (KGE) models rely on precise factual information to learn effective representations. These learned representations support many downstream tasks, with link prediction being a primary application. However, recent studies have shown that noise in training data can compromise the effectiveness of knowledge graph embeddings. Therefore, KGEs are highly vulnerable to data poisoning attacks. Typically current attacks on KGEs are studied under targeted scenarios, where target facts and the model are assumed to be known beforehand. Yet, this information is not often available in real-world scenarios. Thus, more realistic scenarios involve non-targeted but malicious perturbations aimed at reducing the overall model performance. In this paper, we focus on enhancing the robustness of link prediction approaches in non-targeted settings. To mitigate the harmful impact of the noisy data, we explore soft-label loss functions as a strategy for reducing overconfidence in model predictions. We performed a thorough evaluation on six state-of-the-art models and five benchmark datasets, with different noise ratios introduced into each dataset. Our results show that soft labels commonly improve the robustness of KGE models across various noise ratios.