rohitg00--ai-engineering-from-scratch
wehub-resource-sync
103 行
4.0 KiB
JSON
103 行
4.0 KiB
JSON
{
|
|
"lesson": "25-entity-linking",
|
|
"title": "Entity Linking & Disambiguation",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What does entity linking add on top of NER?",
|
|
"options": [
|
|
"Translations",
|
|
"Sentiment polarity",
|
|
"Part-of-speech tags",
|
|
"It maps each detected mention to a unique entry in a knowledge base (Wikidata, Wikipedia, or a domain KB)"
|
|
],
|
|
"correct": 3,
|
|
"explanation": "EL turns a mention into a canonical KB id, disambiguating between same-name entities."
|
|
},
|
|
{
|
|
"stage": "pre",
|
|
"question": "What are the two main subtasks in entity linking?",
|
|
"options": [
|
|
"Tokenization and parsing",
|
|
"POS tagging and lemmatization",
|
|
"Embedding and clustering",
|
|
"Candidate generation (shortlist of plausible KB entries) and disambiguation (pick the right one given context)"
|
|
],
|
|
"correct": 3,
|
|
"explanation": "EL decomposes into proposing candidates then ranking them by contextual fit."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Why must mention recall be reported alongside disambiguation accuracy?",
|
|
"options": [
|
|
"Mention recall is mandatory by GDPR",
|
|
"Recall replaces precision",
|
|
"Disambiguation cannot recover from missing candidates; the pipeline is bounded by candidate-generation recall",
|
|
"Recall is required by spaCy"
|
|
],
|
|
"correct": 2,
|
|
"explanation": "If candidates miss the gold entity, no disambiguator can fix it; recall floors pipeline quality."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "How does GENRE perform entity linking?",
|
|
"options": [
|
|
"Decodes the entity's canonical name token-by-token under constrained decoding over a trie of valid KB ids",
|
|
"By BM25",
|
|
"By computing TF-IDF",
|
|
"By PageRank"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "GENRE generates the canonical KB name with constrained decoding to guarantee a valid id."
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What is NIL handling in entity linking?",
|
|
"options": [
|
|
"Predicting a 'not in KB' label when no candidate is a real match (emerging entities, obscure people)",
|
|
"Returning the entire candidate list",
|
|
"Lowercasing the input",
|
|
"Skipping every mention"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "NIL prediction prevents guessing wrong KB ids for entities the KB does not cover."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Why does popularity bias hurt entity linking in specialized domains?",
|
|
"options": [
|
|
"Models trained on web data over-predict frequent entities (e.g. basketball Jordan over the ML researcher Michael I. Jordan)",
|
|
"It biases toward older entities only",
|
|
"Popularity removes NIL",
|
|
"Popularity helps recall"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "Popularity priors skew predictions away from less-common, domain-specific name-clashes."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "What is a safe LLM-EL pattern in 2026?",
|
|
"options": [
|
|
"Provide a candidate list and use constrained JSON output that the LLM can only choose from valid KB ids",
|
|
"Just prompt 'find the entity'",
|
|
"Free-form generation",
|
|
"Skip candidate generation"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "Constraining the LLM to a valid candidate list prevents made-up KB ids and keeps output queryable."
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Why must NER mention boundaries be exact for entity linking to work?",
|
|
"options": [
|
|
"Boundary errors propagate: 'Bank of America' clipped to 'Bank' surfaces wrong candidates and tanks EL recall",
|
|
"Boundary precision only affects NER metrics, not entity-linking candidate quality",
|
|
"Boundaries change tokenization",
|
|
"Boundaries break Wikipedia lookups"
|
|
],
|
|
"correct": 0,
|
|
"explanation": "Mis-bounded mentions retrieve the wrong alias set, propagating errors into disambiguation."
|
|
}
|
|
]
|
|
}
|