quiz.json

{
  "lesson": "07-pos-tagging-parsing",
  "title": "POS Tagging and Syntactic Parsing",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the goal of POS tagging?",
      "options": [
        "Detect sentiment",
        "Translate the sentence",
        "Extract named entities",
        "Assign a grammatical category (noun, verb, etc.) to each token"
      ],
      "correct": 3,
      "explanation": "POS tagging labels each token with its part of speech."
    },
    {
      "stage": "pre",
      "question": "Which tagset is the default for cross-lingual work?",
      "options": [
        "Penn Treebank",
        "Stanford Dependencies",
        "Universal Dependencies",
        "CoNLL-2003"
      ],
      "correct": 2,
      "explanation": "Universal Dependencies provides a coarser, language-agnostic 17-tag set."
    },
    {
      "stage": "check",
      "question": "What does a bigram HMM POS tagger model?",
      "options": [
        "P(tags) only",
        "Dependency parses",
        "Embedding similarities",
        "P(tag | previous tag) transitions plus P(word | tag) emissions, decoded with Viterbi"
      ],
      "correct": 3,
      "explanation": "Bigram HMM uses tag transitions and word emissions; Viterbi finds the highest-probability sequence."
    },
    {
      "stage": "check",
      "question": "What does the Viterbi algorithm compute for an HMM tagger?",
      "options": [
        "The single highest-probability tag sequence via dynamic programming over the tag lattice",
        "The transition matrix",
        "The marginal probability of each tag",
        "The forward probabilities only"
      ],
      "correct": 0,
      "explanation": "Viterbi finds the argmax sequence with O(n * |T|^2) dynamic programming."
    },
    {
      "stage": "check",
      "question": "What does dependency parsing produce?",
      "options": [
        "A tree where each word has one head word and a labeled grammatical relation",
        "A flat BIO sequence",
        "A constituency tree of NP/VP/PP labels",
        "A coreference chain"
      ],
      "correct": 0,
      "explanation": "Dependency parses give per-word (head, relation) edges; constituency parses give nested phrase structures."
    },
    {
      "stage": "post",
      "question": "Why is the accuracy ceiling on PTB POS tagging around 97-98%?",
      "options": [
        "Models cannot exceed 98% in any task",
        "Universal Dependencies caps at 98%",
        "Human annotators only agree about 97% of the time, so models above ~98% may be overfitting test data",
        "Hardware limitations"
      ],
      "correct": 2,
      "explanation": "Annotator disagreement bounds the achievable accuracy; very high numbers often signal overfitting."
    },
    {
      "stage": "post",
      "question": "Why does POS tagging still matter in 2026 LLM pipelines?",
      "options": [
        "LLMs cannot run without it",
        "Lemmatization, aspect-based sentiment, query decomposition, structured-output validation, and cross-lingual transfer all still consume POS or dependency parses",
        "Required for tokenization",
        "Replaces transformers"
      ],
      "correct": 1,
      "explanation": "POS and dependency parses feed many structured downstream tasks even when LLMs generate the prose."
    },
    {
      "stage": "post",
      "question": "Which library should you reach for in most production POS / parse tasks?",
      "options": [
        "scikit-learn",
        "NumPy",
        "spaCy (or stanza/trankit for top accuracy or wider language coverage)",
        "Roll your own parser"
      ],
      "correct": 2,
      "explanation": "spaCy ships fast production-grade POS + dependency parsers; stanza/trankit cover broader languages."
    }
  ]
}