Named Entity Recognition

Named Entity Recognition (NER) finds entity mentions and assigns labels such as person, location, or disease, helping turn unstructured data into searchable information for applications like knowledge graphs, clinical systems, and privacy protection. Common approaches include fine-tuned transformer encoders with token- or span-level prediction, CRF decoders, and joint extraction models; newer work also uses generative language models, retrieval, and multimodal grounding. Benchmarks and datasets increasingly cover specialized domains, low-resource languages, historical texts, and nested or fine-grained entities, while evaluation often focuses on exact span-level F1 and cross-domain transfer. A central tension is that compact, supervised encoders often remain more accurate and efficient than prompted generative models, while LLMs offer flexible labels and low-resource adaptation but can struggle with boundaries, domain shift, and consistent annotation conventions.