Zero-Shot Named Entity Recognition (NER)

Zero-Shot NER for PII detection using GLiNER, NuNER, and Spacy on Indian, African, Asian, and European names

Image generated by Author using ChatGPT

GLiNER and NuNER are zero-shot Named Entity Recognition (NER) models: You spell out the entity you want to detect, such as “person,” “organization,” “phone number,” etc., and the model will find those entities for you without any model training! NER problems are notoriously complex to solve because of the two issues mentioned below. This article was written entirely by a human with help from Grammarly’s grammar checker, which has been my writing method since 2019.

1. Data labeling is expensive: Human labelers must meticulously annotate every word in the text as a B- the beginning of an entity, I- an intermediate word in an entity, or O- not an entity of interest in the IOB tagging strategy.
2. Model training has difficulty converging: Most models treat NER as a token classification problem where we classify every word or token in the text as the B, I, or O entity tags.

Hence, having a robust zero-shot NER where we don’t have to collect labels or train models massively saves time and money!

NER has many applications, especially in the age of large language models (LLM) hosted by third parties like…