引導式測驗 · 中等 等級
命名實體識別與分類文字標註測驗
將 NER 範圍與文件標籤、標記對齊和重疊實體進行比較,同時選擇適合模型和任務的模式。
問題 1 的 8
Which output distinguishes NER from document classification?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- Which output distinguishes NER from document classification?
- NER marks labeled text spans; classification labels the whole document
- NER assigns one sentiment to a document; classification marks character offsets
- Both tasks only label audio
- Classification always uses token-level BIO labels
- Why does tokenization create a challenge for NER annotation alignment?
- Tokenizers refuse to process any text with entities
- A word like 'COVID-19' may be split into multiple subword tokens, so a human's character-level span must be aligned to those token boundaries
- Tokenization deletes all punctuation automatically
- Tokenizers only work on classification tasks, not NER
- In the BIO tagging scheme, what does the tag 'I-PERSON' indicate?
- The token is the beginning of a person entity
- The token is outside any entity
- The token is inside a person entity that has already begun
- The token belongs to an organization, not a person
- Why can a phrase like 'Bank of America Tower' be difficult to annotate with a simple entity-tagging format?
- It contains no entities at all
- It is too short to tokenize
- It only applies to classification tasks, not NER
- It may require a nested entity, such as a company name inside a building name, which simple formats can't represent
- What does Cohen's kappa measure in the context of text annotation?
- How much agreement two annotators have when labeling the same text, beyond what chance would predict
- The processing speed of the annotation software
- The number of tokens in a document
- The accuracy of a trained model on test data
- What does low inter-annotator agreement usually indicate about a labeling project?
- The annotators need faster computers
- The schema or instructions may be ambiguous, so investigate them alongside annotator calibration
- The dataset is too large
- The model has already been trained incorrectly
- Why is it a misconception to assume document classification always requires less schema work than NER?
- Classification is always technically impossible
- NER never has any ambiguity issues
- Ambiguous document-level categories, like distinguishing 'complaint' from 'feedback', can cause as much disagreement as span-level entity boundaries
- Classification only applies to images, not text
- How do some frameworks handle overlapping entities differently from the standard BIO scheme?
- One document-level sentiment score
- Only a BIO tag per character
- A word-count total
- A separate span with start, end, and label fields, if the model supports it