NLP Named Entity Recognition 2 — Questions and Answers
Question 1: Which tagging scheme uses B- prefix for the beginning of an entity and I- prefix for continuation?
- IOB2 (BIO) (Correct answer)
- BILOU
- IO
- BIOES
Correct answer: IOB2 (BIO)
IOB2 (also called BIO) uses B- for the first token of an entity and I- for subsequent tokens, with O for non-entity tokens.
Question 2: What is the primary advantage of the BILOU tagging scheme over BIO?
- It reduces vocabulary size
- It explicitly marks the last token and unit entities, aiding disambiguation (Correct answer)
- It requires fewer training examples
- It is compatible with all tokenizers
Correct answer: It explicitly marks the last token and unit entities, aiding disambiguation
BILOU adds L- (Last) and U- (Unit/singleton) tags, helping models distinguish entity boundaries more precisely than BIO.
Question 3: In a CRF-based NER model, what does the CRF layer model that a simple softmax classifier does not?
- Word embeddings
- Label dependencies between adjacent tokens (Correct answer)
- Character-level features
- Out-of-vocabulary words
Correct answer: Label dependencies between adjacent tokens
CRF (Conditional Random Field) models the joint probability of the entire label sequence, capturing dependencies between consecutive labels.
Question 4: Which evaluation metric is most commonly reported for NER systems?
- Accuracy
- BLEU score
- Entity-level F1 score (Correct answer)
- Perplexity
Correct answer: Entity-level F1 score
NER is evaluated using entity-level F1, where an entity is counted correct only if both its span and type are exactly correct.
Question 5: What challenge does 'nested NER' address?
- Entities that span multiple sentences
- Entities where one entity span is contained within another (Correct answer)
- Entities with ambiguous types
- Entities in low-resource languages
Correct answer: Entities where one entity span is contained within another
Nested NER handles cases like 'Bank of America' (ORG) containing 'America' (LOC), where standard flat NER misses inner entities.
Question 6: Which dataset is a well-known English NER benchmark that includes PER, ORG, LOC, and MISC entity types from news wire?
- SQuAD
- CoNLL-2003 (Correct answer)
- SNLI
- OntoNotes 5.0
Correct answer: CoNLL-2003
CoNLL-2003 is the canonical English NER benchmark derived from Reuters news, with four entity categories: PER, ORG, LOC, and MISC.
Question 7: What problem does 'mention detection' solve as a subtask before full NER?
- Resolving coreference chains
- Identifying entity span boundaries without classifying type (Correct answer)
- Translating entity names
- Linking entities to a knowledge base
Correct answer: Identifying entity span boundaries without classifying type
Mention detection identifies where entity spans are in text, leaving type classification to a subsequent step, which can improve pipeline modularity.
NLP Named Entity Recognition 2
Which tagging scheme uses B- prefix for the beginning of an entity and I- prefix for continuation?