Rohit Ghumare 836a6f29e8 feat(phase-05/06): named entity recognition ...

BIO tagging scheme with round-trip helpers (spans_to_bio and
bio_to_spans with tested equivalence). Rule-based gazetteer baseline,
hand-crafted feature extraction (word shape, suffix, neighbors) used
by classical CRFs, and sketched BiLSTM-CRF architecture.

Walks the evolution: rule-based (zero OOV coverage) -> CRF (hand
features, enforces valid BIO) -> BiLSTM-CRF (learned features) ->
transformer fine-tune (best accuracy, most compute). Each step names
the limitation its successor solved.

Ship artifact: ner-picker skill that matches approach to domain,
label set, language, latency, and data volume. Rejects transformer
fine-tune under 500 labeled examples without a pretrained domain model.

~75 minutes. Prerequisites lesson 05/02.

2026-04-22 15:46:23 +01:00

..

.gitkeep

feat(phase-05/06): named entity recognition

2026-04-22 15:46:23 +01:00

skill-ner-picker.md

feat(phase-05/06): named entity recognition

2026-04-22 15:46:23 +01:00