Named Entity Recognition (NER) is an important task in natural language processing. However, traditional supervised NER requires large-scale annotated datasets. Distantly supervision is proposed to alleviate the massive demand for datasets, but datasets constructed in this way are extremely noisy and have a serious unlabeled entity problem. The cross entropy (CE) loss function is highly sensitive to unlabeled data, leading to severe performance degradation. As an alternative, we propose a new loss function called NRCES to cope with this problem. A sigmoid term is used to mitigate the negative impact of noise. In addition, we balance the convergence and noise tolerance of the model according to samples and the training process. Experiments on synthetic and real-world datasets demonstrate that our approach shows strong robustness in the case of severe unlabeled entity problem, achieving new state-of-the-art on real-world datasets.