Boundary Detection

The standard token-level annotation scheme for sequence labeling. Each token is tagged as B (Beginning), I (Inside), or O (Outside) of a named entity span.

• B-PER: First token of a person entity
• I-PER: Subsequent tokens of a person entity
• O: Token not part of any entity

Boundary errors occur when a model predicts I-PER without a preceding B-PER, creating an illegal transition that violates the schema's constraints.

A NER paradigm that bypasses token-level BIO tagging entirely. The model directly enumerates all possible text spans up to a maximum length and classifies each as an entity type or null.

• Flat NER: Enumerates spans from a single layer
• Enumeration: Scores all start-end token pairs
• Advantage: Eliminates illegal BIO transitions by design

This approach naturally handles nested entities without the structural limitations of linear-chain sequence models.

The primary evaluation metric for boundary detection quality. A predicted entity is counted as correct only if both its span boundaries and its type classification match the gold standard exactly.

• Precision: Correct predictions / Total predictions
• Recall: Correct predictions / Total gold entities
• F1: Harmonic mean of precision and recall

A boundary-only F1 variant evaluates span detection independently of type accuracy, isolating the model's segmentation performance.

A dynamic programming algorithm that finds the most probable sequence of hidden states in a linear-chain Conditional Random Field (CRF). It efficiently computes the maximum over all possible label paths.

• Input: Emission scores from a neural encoder
• Output: Globally optimal BIO tag sequence
• Constraint: Learns valid transitions (e.g., I-PER must follow B-PER)

Viterbi decoding enforces structural consistency, dramatically reducing illegal boundary predictions compared to independent token classification.

A deep learning layer that applies a bilinear transformation to pairs of input vectors. In span-based NER, it scores the relationship between the start token and end token of a candidate entity.

• Mechanism: score(i,j) = h_i^T U h_j + W[h_i; h_j] + b
• Output: A score for every possible span in the sentence
• Advantage: Models the start-end dependency directly

Used in architectures like Global Pointer to replace CRF layers while maintaining global span coherence.

The benchmark dataset and shared task that established the standard evaluation methodology for language-independent NER. It defines four entity types evaluated with exact boundary matching.

• Entity Types: PER, LOC, ORG, MISC
• Languages: English and German news wire text
• Metric: Mention-level F1 with exact span matching

A boundary detection error on CoNLL-2003—such as predicting "Bank of" instead of "Bank of America"—results in a complete miss, penalizing both precision and recall.