引導式測驗 · 中等 等級

條件隨機場測驗

測試條件序列標記、CRF 與相關模型、訓練和解碼以及結構化輸出評估。

相關引導路徑條件隨機場
問題 1 的 8

What does a CRF estimate in a sequence-labeling task?

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。

  1. What does a CRF estimate in a sequence-labeling task?
    • A single class label without considering neighboring outputs
    • A deterministic rule that never uses training data
    • The probability of an output label sequence conditioned on the observed input
    • The probability of every possible input regardless of labels
  2. How does the original CRF formulation use input features?
    • It can combine multiple overlapping observation features with label-transition features
    • It requires each observation feature to be statistically independent
    • It models only the previous label and ignores the observed input
    • It converts every sequence into an unordered bag of labels
  3. How does a CRF differ from an HMM in the probability it models?
    • A CRF is a neural network and an HMM is a database
    • An HMM always uses word labels but a CRF only handles images
    • A CRF predicts without training labels while an HMM requires none
    • A CRF directly models labels conditioned on observations, while an HMM models a joint process over observations and states
  4. Why can a CRF help with named-entity spans?
    • It uses a separate independent classifier for each word only
    • It scores a label sequence while accounting for neighboring label compatibility
    • It guarantees every entity name is factually correct
    • It detects entities without input features or labeled data
  5. What limitation of locally normalized directed models did the original CRF paper address?
    • A shortage of unlabeled text at inference time
    • A need to translate every observation into an image
    • Label bias toward states with fewer successor states
    • A requirement that inputs have no capitalization
  6. What does a partition function do in a CRF?
    • Converts a sequence model into an HMM
    • Normalizes scores across possible output sequences to form a conditional probability distribution
    • Chooses the number of training examples
    • Removes the need to define output labels
  7. Which evaluation is appropriate for a named-entity CRF?
    • Inspect complete-span errors and report sequence-relevant precision or recall
    • Report only training accuracy on individual tokens
    • Count the number of features and treat it as model quality
    • Check only whether one label appears in the vocabulary
  8. When might a CRF be a poor fit without further modeling?
    • When output labels depend on adjacent labels
    • When labeled sequences are available
    • When the input contains multiple features
    • When the task has important long-range or graph-shaped dependencies beyond the chosen chain structure