Out-of-vocabulary words can be challenging for NER systems. We introduce a self-supervised system for Named Entity Recognition based on a few-shot annotated examples provided by experts. Subsequently, the rest of the words are tagged using the closest similarity match between the word embeddings of each category, generated in the same context as the original annotations. Additionally, we use a dual-threshold scheme to improve the robustness of the method. Our results show that this method outperforms current state-of-the-art methods in both accuracy and generalisation.