Entity Extraction

The foundational task of locating and classifying named entities in text into predefined categories. Modern systems use transformer-based models fine-tuned on annotated corpora to identify spans and assign labels.

• Common entity types: PERSON, ORG, GPE, DATE, MONEY, PERCENT
• Architecture: Token classification heads on top of BERT or similar encoders
• Example: "Apple acquired Beats for $3B" → [Apple:ORG] acquired [Beats:ORG] for [$3B:MONEY]
• Challenge: Handling nested entities like "University of California, Berkeley" where multiple entity spans overlap

The process of connecting a textual mention to its unique canonical entry in a knowledge base. This resolves ambiguity by mapping surface forms like "Washington" to the correct entity—state, person, or city—based on context.

• Knowledge bases: Wikidata, DBpedia, proprietary enterprise graphs
• Key technique: Candidate generation followed by cross-encoder re-ranking
• Example: "Tesla stock surged" → links to wd:Q478214 (the company), not wd:Q355 (the inventor)
• Critical for: Factual grounding and preventing hallucination in RAG pipelines

Identifies all expressions in a text that refer to the same real-world entity, linking pronouns and nominal phrases to their antecedents. Essential for maintaining coherent entity tracking across multi-sentence contexts.

• Span-based models: Predict antecedent links between all mention spans
• Example: "Alice went to the store. She bought milk." → She resolves to Alice
• Enterprise value: Ensures complete entity extraction from legal contracts and medical records
• Integration: Often paired with NER to cluster all mentions of the same extracted entity

Extends traditional NER by assigning hierarchical, ultra-specific type labels beyond coarse categories. Instead of just PERSON, a mention might be typed as [actor, activist, Nobel laureate].

• Type systems: FIGER (112 types), Ultra-Fine (10,000+ types)
• Architecture: Uses label embeddings and mention-context attention
• Example: "Elon Musk" → {entrepreneur, CEO, engineer, founder, billionaire}
• Benefit: Enables precise query scoping and richer knowledge graph construction

Identifies semantic relationships between extracted entities within a text. Transforms unstructured sentences into structured triples of the form (subject, predicate, object) for knowledge graph population.

• Approaches: Joint entity-relation models, table-filling, generative seq2seq
• Common relations: founded_by, headquartered_in, acquired, spouse_of
• Example: "Google, based in Mountain View, was founded by Larry Page" → (Google, headquartered_in, Mountain View), (Google, founded_by, Larry Page)
• Output: Directly feeds into enterprise knowledge graphs for deterministic reasoning

Extends entity extraction beyond text to align mentions across modalities—linking a textual name to an image region, a spoken utterance, or a video segment. Critical for unified knowledge representation.

• Technique: Joint embedding spaces mapping text, image patches, and audio to shared entity vectors
• Example: A photo caption "Barack Obama at the podium" grounds the text entity to the bounding box of the person in the image
• Application: Powers multi-modal search in e-commerce catalogs and media archives
• Architecture: Often uses contrastive learning between text and visual entity representations

Extracts specific attributes from a query to populate a predefined template, critical for task-oriented dialogue systems.

• Template: A frame with slots like origin, destination, date for a flight booking query.
• Mechanism: Uses a combination of NER and intent classification to identify and extract the required parameters.
• Example: 'Book a flight to Berlin on Friday' fills destination=Berlin and date=Friday.

Converts a natural language query into a structured, machine-executable logical form representing its meaning.

• Output: SQL, SPARQL, Lambda Calculus, or a custom domain-specific language.
• Execution: The logical form is executed directly against a database or knowledge graph for precise retrieval.
• Example: 'Employees hired after 2020' → SELECT name FROM employees WHERE hire_date > '2020-01-01'

Identifies all expressions in a text that refer to the same real-world entity, linking pronouns to their antecedents.

• Anaphora: Resolving 'it', 'he', 'she' back to the original entity mention.
• Importance: Critical for multi-turn conversations where a user says 'Tell me more about it' after a previous query about a specific entity.
• Example: 'Sundar Pichai is the CEO of Google. He joined in 2004.' → links 'He' to 'Sundar Pichai'.

Analyzes a query to determine its domain, temporal range, or other constraints, effectively narrowing the search space.

• Domain Classification: Routes a query to a specific index (e.g., 'legal', 'medical', 'general').
• Temporal Scoping: Extracts date ranges to filter documents by recency.
• Mechanism: Often uses a combination of entity extraction and a lightweight classifier to set metadata filters before retrieval.