rohitg00--ai-engineering-from-scratch
wehub-resource-sync
103 行
4.0 KiB
JSON
103 行
4.0 KiB
JSON
{
|
|
"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)."
|
|
}
|
|
]
|
|
}
|