Named Entity Recognition (NER)

Named Entity Recognition (NER) is an NLP task that identifies and classifies named entities (e.g., person, organization, location) in text. For local AI operators, NER models extract structured data from documents, enabling downstream tasks like search, redaction, or knowledge graph building. NER is typically a sequence labeling problem: each token gets a label (e.g., B-PER, I-PER, O). Operators run NER via Hugging Face pipelines or specialized models like BERT-based NER, which require modest VRAM (2-4 GB for base models).

NER models often use a transformer encoder (e.g., BERT) with a token classification head. The model outputs a probability distribution over entity tags for each token. Common tag schemes include BIO (Begin, Inside, Outside) or BILOU (Begin, Inside, Last, Outside, Unit). Operators can fine-tune a pre-trained model on custom entity types using Hugging Face Transformers. Inference speed depends on sequence length and hardware: a BERT-base NER model processes ~100 tokens in 10-20 ms on a modern GPU. Quantization (e.g., ONNX or bitsandbytes) reduces VRAM usage by 2-4x with minor accuracy loss. For long documents, operators may split text into chunks with overlap to avoid context window limits.

An operator runs a BERT-based NER model on an RTX 3060 (12 GB VRAM) to extract names and dates from legal contracts. Using Hugging Face's pipeline('ner', model='dslim/bert-base-NER'), the model processes a 512-token chunk in 15 ms. With a batch size of 8, throughput reaches ~50 chunks/sec. The operator must ensure VRAM fits the model (1.5 GB for BERT-base) plus batch overhead.

In a local AI workflow, an operator loads a NER model via Hugging Face Transformers: from transformers import pipeline; ner = pipeline('ner', model='dslim/bert-base-NER', device=0). They then pass text chunks to ner(text) and receive a list of entities with scores. For larger documents, they use tokenizer.batch_encode_plus with padding and truncation. If VRAM is tight, they enable 8-bit quantization via bitsandbytes or switch to a smaller model like distilbert-base-NER.