{
  "lesson": "12-text-summarization",
  "title": "Text Summarization",
  "questions": [
    {
      "stage": "pre",
      "question": "What is the key behavioral difference between extractive and abstractive summarization?",
      "options": [
        "Extractive uses TF-IDF; abstractive uses Word2Vec",
        "Extractive returns sentences verbatim from the source; abstractive generates new text and can hallucinate",
        "Extractive is multilingual only",
        "Extractive is slower than abstractive"
      ],
      "correct": 1,
      "explanation": "Extractive lifts sentences verbatim; abstractive rewrites and risks hallucination."
    },
    {
      "stage": "pre",
      "question": "What does ROUGE measure?",
      "options": [
        "Embedding similarity",
        "N-gram and longest-common-subsequence overlap between system and reference summaries",
        "Reading time",
        "Token-level perplexity"
      ],
      "correct": 1,
      "explanation": "ROUGE-1/2/L measure unigram, bigram, and LCS overlap with references."
    },
    {
      "stage": "check",
      "question": "How does TextRank score sentences in extractive summarization?",
      "options": [
        "By raw word count",
        "By comparing to a reference summary",
        "By running a PageRank-style iteration over a graph where edges are sentence-similarity weights",
        "By embedding cosine to the question"
      ],
      "correct": 2,
      "explanation": "TextRank uses PageRank over a sentence-similarity graph; highly connected sentences score highest."
    },
    {
      "stage": "check",
      "question": "Why enable stemming when computing ROUGE?",
      "options": [
        "Without stemming, 'running' and 'run' count as different tokens and ROUGE undercounts true overlap",
        "To speed up ROUGE",
        "Stemming normalizes case",
        "Stemming is required by the rouge-score package"
      ],
      "correct": 0,
      "explanation": "Stemming merges morphological variants so ROUGE credits semantically equivalent forms."
    },
    {
      "stage": "check",
      "question": "Which 2026 metric is purpose-built to detect summary hallucinations via NLI entailment?",
      "options": [
        "BLEU",
        "ROUGE-L",
        "Faithfulness checks (e.g. FactCC or RAGAS faithfulness) using NLI between source and summary claims",
        "chrF"
      ],
      "correct": 2,
      "explanation": "NLI-based faithfulness scoring flags claims in the summary not entailed by the source."
    },
    {
      "stage": "post",
      "question": "Why is extractive summarization preferred for compliance-adjacent content?",
      "options": [
        "Outputs are lifted verbatim from the source, eliminating the abstractive hallucination class",
        "It is faster",
        "Extractive supports longer outputs",
        "ROUGE scores are higher"
      ],
      "correct": 0,
      "explanation": "Verbatim extraction cannot invent content, which matters where factuality is regulated."
    },
    {
      "stage": "post",
      "question": "Which of these is an abstractive hallucination type to monitor for?",
      "options": [
        "Punctuation drift",
        "Stopword removal",
        "Long sentences",
        "Entity swap (e.g. 'John Smith' rendered as 'John Brown'), number drift, polarity flip, or fact invention"
      ],
      "correct": 3,
      "explanation": "Entity swaps, numeric drift, polarity flips, and invented facts are the canonical abstractive failure modes."
    },
    {
      "stage": "post",
      "question": "When would you reach for a Pegasus checkpoint over BART-large-CNN?",
      "options": [
        "When evaluating BLEU",
        "When the input is short",
        "When you need extractive output",
        "For domains like scientific abstracts where Pegasus's gap-sentence pretraining objective is a closer fit"
      ],
      "correct": 3,
      "explanation": "Pegasus's gap-sentence objective excels at long-form domain summarization (e.g. pubmed)."
    }
  ]
}