{
  "lesson": "06-named-entity-recognition",
  "title": "Named Entity Recognition",
  "questions": [
    {
      "stage": "pre",
      "question": "What is BIO tagging?",
      "options": [
        "Per-token labels: B-TYPE for entity start, I-TYPE for inside, O for outside",
        "A binary entity vs non-entity scheme",
        "A tokenization style",
        "A tree representation of entities"
      ],
      "correct": 0,
      "explanation": "BIO turns span extraction into token classification with B/I/O prefixes."
    },
    {
      "stage": "pre",
      "question": "Why are rule-based gazetteers brittle in production NER?",
      "options": [
        "They require GPUs",
        "They are slow",
        "They have zero coverage on new entities and cannot disambiguate (e.g. Apple fruit vs company)",
        "They cannot handle multi-token entities"
      ],
      "correct": 2,
      "explanation": "Gazetteers can match known strings but cannot disambiguate sense or generalize to unseen names."
    },
    {
      "stage": "check",
      "question": "What is the key advantage of a CRF over an HMM for NER?",
      "options": [
        "CRFs avoid the Viterbi algorithm",
        "CRFs never need training data",
        "CRFs are discriminative and can mix arbitrary features (shape, capitalization, neighbors)",
        "CRFs are faster"
      ],
      "correct": 2,
      "explanation": "CRFs are discriminative and let you condition on rich, overlapping features."
    },
    {
      "stage": "check",
      "question": "In a BiLSTM-CRF architecture, what role does the CRF layer play?",
      "options": [
        "Performs tokenization",
        "Enforces valid BIO tag sequences by modeling tag-to-tag transitions on top of LSTM emissions",
        "Pretrains the LSTM",
        "Replaces embeddings"
      ],
      "correct": 1,
      "explanation": "The CRF on top of LSTM features models inter-label dependencies and rules out illegal sequences."
    },
    {
      "stage": "check",
      "question": "Why must NER be evaluated with entity-level F1, not token-level F1?",
      "options": [
        "Entity-level F1 is easier to compute",
        "Predicted spans must match true spans exactly; token-level F1 overstates accuracy by counting partial matches",
        "Token-level F1 cannot handle BIO",
        "Entity-level F1 is required by HuggingFace"
      ],
      "correct": 1,
      "explanation": "Span exact-match is the meaningful metric; token-level F1 inflates scores via partial overlap."
    },
    {
      "stage": "post",
      "question": "What does aggregation_strategy='simple' do in the HuggingFace NER pipeline?",
      "options": [
        "Lowercases the output",
        "Skips tokenization",
        "Merges contiguous B-X and I-X tokens into a single span",
        "Returns only the most confident entity"
      ],
      "correct": 2,
      "explanation": "It merges contiguous BIO tokens of the same type into span-level entities."
    },
    {
      "stage": "post",
      "question": "Why does standard BIO fail on nested entities?",
      "options": [
        "BIO requires a transformer",
        "BIO is a flat per-token scheme and cannot express two overlapping spans of different types",
        "BIO cannot represent multi-token entities",
        "BIO drops the type label"
      ],
      "correct": 1,
      "explanation": "BIO assigns one label per token; nested spans need multi-pass or span-based models."
    },
    {
      "stage": "post",
      "question": "When does classical NER (CRF or BiLSTM-CRF) still beat an LLM in 2026?",
      "options": [
        "On nested entities",
        "Whenever the input is in English",
        "On open-domain narrative",
        "Under tight latency budgets, abundant labels, stable ontologies, or non-generative regulatory constraints"
      ],
      "correct": 3,
      "explanation": "Classical NER wins on latency, labeled-data regimes, fixed ontologies, and on-prem constraints."
    }
  ]
}