Thực thể được đặt tên Rec. Câu đố

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

1. Before annotating a corpus, what should a team define?
   • Entity categories and span-boundary rules
   • The model vendor’s pricing plan
   • The number of pages in each file
   • A list of every person’s biography
2. Why might “Labs” be included in “Northstar Research Labs” for one dataset?
   • All suffixes are always excluded
   • The annotation rules define whether organization suffixes are part of the span
   • NER decides company ownership
   • Tokenization makes boundaries irrelevant
3. When token labels are aligned to subword tokens, what requires care?
   • Translating every token into a new language
   • Replacing offsets with a summary
   • Mapping labels back to the original text spans
   • Treating all subwords as separate people
4. Why preserve character offsets for an extracted entity?
   • The system can infer a person’s identity
   • Offsets supply missing context automatically
   • They turn NER into sentiment analysis
   • An application can highlight the exact source span
5. Which evaluation check matters for NER?
   • Check only whether a name appears somewhere
   • Assess both entity boundaries and entity types
   • Measure average sentence length only
   • Count model parameters
6. A recognizer tags “Jordan” as a person. What would be needed to link it to a particular real-world record?
   • A longer NER label
   • Another punctuation mark
   • Separate entity linking with additional evidence
   • The same classifier run twice
7. The system returns “Northstar” but the annotated organization span is “Northstar Research Labs.” What kind of error may this be?
   • A span-boundary error even if the type is right
   • Proof that entity linking succeeded
   • A sentiment classification error
   • A change in the source document
8. Why is high average NER performance not enough for redaction?
   • Redaction does not use extracted spans
   • Every NER label hides the whole document
   • Offsets prevent misses
   • A missed sensitive name can remain visible despite a strong average score