This is why rule-based NER is brittle: 'Great' is capitalized because it starts a sentence after a period — not because it's a named entity. The rule 'capitalized word = entity' produces massive false positives at sentence boundaries. CRF and neural models learn to distinguish sentence-initial capitalization from entity capitalization by also considering: (1) the word's frequency, (2) whether it appears capitalized mid-sentence elsewhere, (3) surrounding context. This single example shows why ML-based NER replaced rules.

Without a CRF layer, each token is labeled independently. The model sees 'New' (capitalized, location-like) → B-LOC. Then 'York' (capitalized, location-like) → B-LOC again, starting a NEW entity instead of continuing. 'City' (common word) → O. This fragments 'New York City' into two entities: 'New' and 'York.' A CRF adds transition scores: P(I-LOC | B-LOC) >> P(B-LOC | B-LOC), so the optimal sequence becomes B-LOC, I-LOC, I-LOC — correctly recognizing the full 3-word entity.

This is exactly what contextual embeddings solve. In a non-contextual model (Word2Vec), 'apple' always has the SAME vector regardless of context. But BERT produces DIFFERENT embeddings for each occurrence: 'ate an apple' context → fruit → O tag. 'reading about Apple' context → company → B-ORG. 'Apple Watch' context → brand + product → B-ORG I-ORG. This context-dependent representation is why BERT revolutionized NER.

Entity ambiguity is one of NER's hardest problems. 'Washington' can be a person, state, city, or university. 'Amazon' can be company, river, or streaming service. NER labels span + type, but entity linking (disambiguation) resolves WHICH entity. Modern systems use a pipeline: NER extracts mentions → entity linking matches each to a knowledge base using context. Wikipedia-based entity linking achieves ~90% accuracy by encoding context and candidate descriptions with cross-encoders.