quiz.json

{
  "lesson": "24-coreference-resolution",
  "title": "Coreference Resolution",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the goal of coreference resolution?",
      "options": [
        "Cluster all mentions (named, nominal, pronominal) that refer to the same real-world entity",
        "Tag parts of speech",
        "Extract relations",
        "Translate pronouns to nouns"
      ],
      "correct": 0,
      "explanation": "Coref clusters mention spans that all refer to the same entity."
    },
    {
      "stage": "pre",
      "question": "Which type of expression is a 'nominal' mention?",
      "options": [
        "A proper noun only",
        "A noun phrase such as 'the CEO' or 'the company'",
        "A pronoun like 'she'",
        "A verb"
      ],
      "correct": 1,
      "explanation": "Nominal mentions are noun phrases like 'the company'; pronominal mentions are pronouns."
    },
    {
      "stage": "check",
      "question": "What is the modern (Lee et al., 2017) coref architecture?",
      "options": [
        "BM25 retrieval",
        "Rule-based syntactic parsing only",
        "Decision trees",
        "End-to-end span-based: enumerate spans, score mentions, then score antecedent probabilities and cluster greedily"
      ],
      "correct": 3,
      "explanation": "End-to-end neural coref enumerates spans and learns mention + antecedent scoring jointly."
    },
    {
      "stage": "check",
      "question": "What does CoNLL F1 average?",
      "options": [
        "F1 across languages",
        "MUC, B-cubed, and CEAF-phi4 F1 scores",
        "Token and span F1",
        "Precision and recall"
      ],
      "correct": 1,
      "explanation": "CoNLL F1 is the mean of MUC, B-cubed, and CEAF-phi4 metrics."
    },
    {
      "stage": "check",
      "question": "What is bridging anaphora?",
      "options": [
        "A mistranslation",
        "An implicit reference like 'the wheels' implying the wheels of a previously mentioned car",
        "A pronoun before its referent",
        "A pronoun without an antecedent"
      ],
      "correct": 1,
      "explanation": "Bridging links a mention to a part-of or related entity that was implied but not explicitly stated."
    },
    {
      "stage": "post",
      "question": "Why is LLM-only coref unreliable on long documents?",
      "options": [
        "LLMs cannot read text",
        "Tokenizers fail",
        "Single-call LLMs over-merge or silently drop mentions across 50+ paragraphs; require sliding-window plus merge",
        "Coref requires a CFG"
      ],
      "correct": 2,
      "explanation": "Long-doc LLM coref degrades; sliding-window with cross-window merging mitigates."
    },
    {
      "stage": "post",
      "question": "Why merge coref clusters into NER results before downstream tasks?",
      "options": [
        "Required by Wikidata",
        "To increase token count",
        "Lower latency",
        "So downstream tasks see one entity per cluster rather than one per surface mention, dramatically improving coverage"
      ],
      "correct": 3,
      "explanation": "Without merging, NER counts each surface form separately and misses 60-80% of entity mentions."
    },
    {
      "stage": "post",
      "question": "Why are hard-coded gender rules a fragility in coref systems?",
      "options": [
        "They break on non-binary referents, organizations, and animals; learned scoring is more robust",
        "They require GPU",
        "They run too fast",
        "They cannot use POS tags"
      ],
      "correct": 0,
      "explanation": "Gender heuristics fail in demographically diverse text; learned models are preferred."
    }
  ]
}