Nested Ner
Nested NER handles entities within entities — recognizing that "Bank of America" contains both an organization ("Bank of America") and a location ("America"), or that "New York University Medical Center" has nested organization and location entities.
What Is Nested NER?
- Definition: Recognize overlapping or nested entity mentions.
- Example: "Bank of [America]LOC" is also "[Bank of America]ORG".
- Challenge: Traditional NER assumes non-overlapping entities.
Nested Entity Examples
Organization + Location: "Bank of [America]LOC" → "[Bank of America]ORG".
Person + Organization: "[Michael]PER [Jordan]PER" → "[Michael Jordan]PER".
Product + Organization: "[Microsoft]ORG [Windows]PRODUCT" → "[Microsoft Windows]PRODUCT".
Location Hierarchy: "[New York]CITY [City]" → "[New York City]CITY".
Why Nested NER?
- Completeness: Capture all entity mentions, not just outermost.
- Precision: Distinguish "America" (location) from "Bank of America" (organization).
- Knowledge Extraction: Build richer knowledge graphs.
- Domain-Specific: Medical, legal texts have complex nested entities.
Approaches
Layered Tagging: Multiple NER passes for different nesting levels.
Span-Based: Enumerate all possible spans, classify each.
Hypergraph: Model nested structure as hypergraph.
Transition-Based: Parse entities like syntactic parsing.
Neural Models: Span-based BERT models, nested attention.
Challenges: Exponential span candidates, ambiguous boundaries, rare nested patterns, computational cost.
Applications: Biomedical NER (nested gene/protein names), legal documents, news analysis, knowledge base construction.
Tools: Nested NER models in research, spaCy with custom components, specialized biomedical NER systems.