Fine-Grained Entity Typing
Fine-grained entity typing classifies entities into detailed, specific types — going beyond coarse categories (person, organization, location) to fine-grained types like "politician," "software company," "mountain," enabling more precise entity understanding and knowledge extraction.
What Is Fine-Grained Entity Typing?
- Definition: Classify entities into specific, detailed types.
- Coarse: PERSON, ORGANIZATION, LOCATION (3-10 types).
- Fine-Grained: politician, athlete, actor, software_company, mountain, river (100-10,000 types).
Type Hierarchies
PERSON → politician, athlete, actor, scientist, musician, author.
ORGANIZATION → company, university, government_agency, non_profit.
LOCATION → city, country, mountain, river, building, landmark.
PRODUCT → software, vehicle, food, drug, weapon.
EVENT → war, election, natural_disaster, sports_event.
Why Fine-Grained Types?
- Precision: "Apple" as "technology_company" vs. "fruit".
- Knowledge Graphs: Richer entity representations.
- Question Answering: "Which politician...?" — need to identify politicians.
- Relation Extraction: Type constraints on relations (CEOs lead companies).
- Search: Filter by specific entity types.
Challenges
Type Ambiguity: Entities can have multiple types (Obama: politician, author, lawyer).
Type Granularity: How specific should types be?
Rare Types: Long-tail types with few training examples.
Type Hierarchy: Manage hierarchical type relationships.
Scalability: Thousands of types vs. traditional 3-10 types.
Approaches
Multi-Label Classification: Assign multiple types per entity.
Hierarchical Classification: Leverage type hierarchy.
Zero-Shot: Classify into types not seen during training.
Distant Supervision: Use knowledge bases for training labels.
Neural Models: BERT-based fine-grained typing.
Applications: Knowledge base construction, question answering, information retrieval, semantic search, relation extraction.
Datasets: FIGER, OntoNotes, BBN, Ultra-Fine Entity Typing.
Tools: Research systems, custom fine-grained typing models, knowledge base APIs (Wikidata, DBpedia).