NLP Named Entity Recognition 3 — Questions and Answers
Question 1: Which character-level feature is commonly added to NER models to help identify proper nouns?
- Part-of-speech tags
- Capitalization features (Correct answer)
- Dependency parse labels
- Sentence length
Correct answer: Capitalization features
Capitalization is a strong surface cue for proper nouns (entities) in English and many other languages, so it is a standard NER feature.
Question 2: In transformer-based NER, why is subword tokenization (e.g., WordPiece) potentially problematic?
- It increases model size
- Entity boundaries may fall within a subword split, complicating label alignment (Correct answer)
- It prevents the use of CRF layers
- It removes positional encodings
Correct answer: Entity boundaries may fall within a subword split, complicating label alignment
Subword tokenization can split a single word into multiple tokens, requiring strategies to align word-level NER labels with subword tokens.
Question 3: What is 'gazetteer' in the context of NER?
- A neural attention mechanism
- A list of known entities used as a lookup feature (Correct answer)
- A type of recurrent cell
- A regularization technique
Correct answer: A list of known entities used as a lookup feature
A gazetteer is a dictionary of known entity names (e.g., cities, companies) whose presence in text provides strong evidence for NER models.
Question 4: Which NER approach does NOT require annotated training data?
- BiLSTM-CRF
- Dictionary-based / rule-based NER (Correct answer)
- Fine-tuned BERT
- CRF with hand-crafted features
Correct answer: Dictionary-based / rule-based NER
Dictionary-based and rule-based NER systems rely on predefined lists and patterns rather than labeled examples, requiring no annotated corpus.
Question 5: What is 'entity linking' (also called entity disambiguation) in relation to NER?
- Connecting entities across coreference chains in a document
- Mapping a recognized entity mention to a specific entry in a knowledge base (Correct answer)
- Linking entity tokens to their POS tags
- Merging overlapping entity spans
Correct answer: Mapping a recognized entity mention to a specific entry in a knowledge base
Entity linking maps surface mentions (e.g., 'Apple') identified by NER to canonical knowledge base entries (e.g., Apple Inc. vs. the fruit).
Question 6: Which of the following best describes the 'O' label in BIO tagging?
- The onset of a new sentence
- A token that belongs to no named entity (Correct answer)
- An optional entity type
- An out-of-vocabulary token
Correct answer: A token that belongs to no named entity
In BIO tagging, O (Outside) marks tokens that are not part of any named entity, forming the majority class in most NER datasets.
Question 7: What technique is used to handle the class imbalance problem in NER, where O labels vastly outnumber entity labels?
- Oversampling entity sentences
- Focal loss or weighted cross-entropy (Correct answer)
- Removing O-labeled tokens from training
- Using only entity tokens for backpropagation
Correct answer: Focal loss or weighted cross-entropy
Focal loss down-weights easy (O-label) examples and focal cross-entropy assigns higher loss weight to rare entity classes to counter imbalance.
NLP Named Entity Recognition 3
Which character-level feature is commonly added to NER models to help identify proper nouns?