NLP Named Entity Recognition 4 — Questions and Answers
Question 1: Which pre-trained model popularized the use of contextual embeddings for NER and achieved state-of-the-art results on CoNLL-2003?
- Word2Vec
- GloVe
- BERT (Correct answer)
- FastText
Correct answer: BERT
BERT's contextual embeddings, fine-tuned with a token classification head, set new state-of-the-art records on CoNLL-2003 NER in 2018.
Question 2: What is 'zero-shot NER'?
- NER on text with no punctuation
- Recognizing entity types not seen during training (Correct answer)
- NER without a tokenizer
- Predicting entities at zero latency
Correct answer: Recognizing entity types not seen during training
Zero-shot NER recognizes entity types absent from the training set, typically by leveraging type descriptions or prompting large language models.
Question 3: In few-shot NER, what is a 'support set'?
- The full training corpus
- A small set of labeled examples provided at inference time for new entity types (Correct answer)
- A list of stop words to filter
- A validation set for hyperparameter tuning
Correct answer: A small set of labeled examples provided at inference time for new entity types
In few-shot NER, the support set contains a handful of labeled examples per new entity type that the model uses to generalize at inference.
Question 4: Which domain is known for particularly challenging NER due to highly specialized terminology and non-standard abbreviations?
- Sports news
- Biomedical / clinical text (Correct answer)
- Weather reports
- Social media memes
Correct answer: Biomedical / clinical text
Biomedical NER must recognize genes, proteins, diseases, and drugs with complex nomenclature, making it harder than general-domain NER.
Question 5: What is 'cross-lingual NER'?
- NER that switches languages mid-sentence
- Training on one language and transferring the model to recognize entities in another (Correct answer)
- NER that merges entities across parallel corpora
- Using multiple NER models in an ensemble
Correct answer: Training on one language and transferring the model to recognize entities in another
Cross-lingual NER trains on a high-resource language and transfers to a low-resource target language, leveraging multilingual embeddings like mBERT or XLM-R.
Question 6: Why is NER particularly difficult on Twitter/social media data compared to newswire?
- Social media posts are always longer
- Informal spelling, abbreviations, slang, and unconventional capitalization violate standard NER assumptions (Correct answer)
- Social media has no named entities
- Twitter enforces strict grammatical structure
Correct answer: Informal spelling, abbreviations, slang, and unconventional capitalization violate standard NER assumptions
Social media text features inconsistent capitalization, abbreviations, hashtags, and misspellings that break lexical features and pre-trained models trained on formal text.
Question 7: Which span-based NER approach enumerates all candidate spans and classifies each one, enabling nested entity detection?
- Sequence labeling with BIO tags
- Span classification models (Correct answer)
- Hidden Markov Models
- Rule-based matching
Correct answer: Span classification models
Span-based models score all possible text spans and classify each independently, naturally handling nested entities that sequential BIO tagging cannot represent.
NLP Named Entity Recognition 4
Which pre-trained model popularized the use of contextual embeddings for NER and achieved state-of-the-art results on CoNLL-2003?