Named Entity Recognition (NER) plays a vital role in extracting specialized entities from text, particularly in biomedical and chemical domains. Building reliable NER models in these fields is challenging due to the scarcity of annotated data, with manual annotation being a primary method for data creation. However, manual annotation is both time-intensive and requires domain expertise, making it impractical for specialized domains. Existing data augmentation (DA) techniques also depend on annotated data to some degree, limiting their applicability. To address this, we introduce a novel synthetic data generation approach for NER that leverages large language models (LLMs) to generate sentences using only a set of example entities. This approach streamlines the augmentation process and remains effective even with a minimal number of entities. We assess its effectiveness on the BC4CHEMD and BC5CDR datasets using BERT-based models, demonstrating that synthetic data enhances model performance and robustness, particularly in low-resource scenarios. Our work offers a scalable solution for improving NER in specialized domains, mitigating the constraints of manual annotation and conventional augmentation methods.