Foundations
Named Entity Recognition (NER)
Named Entity Recognition is an AI technique that scans text and automatically identifies and labels specific real-world items in it, such as people, organizations, dates, or medications.
The Core Idea
Most business documents contain a mix of free-flowing prose and specific factual items buried inside it — a patient's name, a drug dosage, an invoice number, a company mentioned in a contract. Named Entity Recognition, or NER, is the technique that finds and labels those items automatically, tagging spans of text as belonging to categories like person, organization, date, medication, or monetary amount. It's one of the older and more mature techniques in the NLP toolkit, predating today's large language models by decades, though modern LLM-based systems now often perform this extraction as one step within a broader task rather than as a standalone classifier. The output is structured data, a list of typed entities with their positions in the text, which is what makes NER genuinely useful downstream: unstructured prose becomes something a database, a form, or a business rule engine can act on directly.
Beyond Simple Keyword Matching
Good NER isn't a lookup table of known names — it has to handle ambiguity from context, since the same word can be a different entity type depending on where and how it appears, like a place name that's also a common person's name, or a drug name that's also a common English word. Domain matters enormously: an NER model tuned for general news text will miss or misclassify clinical dosages, legal party names, or Arabic honorifics and naming conventions unless it's been trained or adapted on text from that specific domain. This is why off-the-shelf NER tools often need retraining or fine-tuning with domain-specific examples before they're reliable for specialized business use, and why accuracy claims for a general-purpose NER model don't transfer automatically to a specialized use case. Evaluation also has to be entity-type specific — a system might extract dates near-perfectly while badly under-extracting medication names, and averaging those together into one accuracy number hides the gap that actually matters operationally.
Where It Fits at Hayya Med AI
NER is a core extraction step in Hayya Med AI's document pipelines — pulling patient identifiers, medications, dosages, and diagnosis codes out of free-text clinical notes so they can populate structured EHR fields, and separately, pulling supplier names, line-item amounts, and contract dates out of procurement invoices and tenders submitted in a mix of Arabic and English. Because both domains involve entities that general-purpose NER models handle poorly out of the box, such as Arabic medication brand names and Qatari company registration formats, the extraction models are adapted with domain-specific training examples rather than deployed as-is, and entity-level accuracy, not an averaged score, is what the team tracks before trusting output to flow into a billing or compliance system.
Written by Abbas Al Masri
Founder & Chief Executive Officer, Hayya Med AI
Abbas Al Masri founded Hayya Med AI to help organizations across the GCC and beyond build AI-native platforms grounded in real market, regulatory, and operational reality.
View Full Profile →Frequently Asked
Is NER the same as OCR?
No — OCR converts an image of text into machine-readable text; NER then finds and labels specific facts within that text, so the two are typically used together in document-processing pipelines but solve different problems.
Can NER work reliably across both Arabic and English in the same document?
Only if the underlying model was trained or adapted on bilingual and code-switched text; a model trained purely on English struggles with Arabic script, name structures, and mixed-language sentences common in Gulf business documents.