entity linking

**Entity linking** (also called **entity resolution** or **named entity disambiguation**) is the NLP task of identifying mentions of entities in text and connecting them to corresponding entries in a **knowledge base** (like Wikipedia, Wikidata, or a domain-specific ontology). It bridges the gap between unstructured text and structured knowledge. **How Entity Linking Works** - **Step 1 — Mention Detection**: Identify spans of text that refer to entities (e.g., "Apple" in "Apple released a new phone"). - **Step 2 — Candidate Generation**: Generate a list of possible knowledge base entries the mention could refer to (Apple Inc., apple fruit, Apple Records, etc.). - **Step 3 — Disambiguation**: Use context to select the correct entity. "Apple released a new phone" → **Apple Inc.** vs. "I ate an apple" → **the fruit**. **Why Entity Linking Matters for RAG** - **Grounding**: Links free-text queries and documents to **canonical entities**, enabling structured reasoning about entities and their relationships. - **Knowledge Graph Integration**: Once entities are linked, you can traverse a **knowledge graph** to find related entities, properties, and facts. - **Disambiguation**: Resolves ambiguity — "Python" could mean the programming language, the snake, or Monty Python depending on context. - **Cross-Document Coreference**: Recognizes that "TSMC," "Taiwan Semiconductor," and "the Taiwanese chipmaker" all refer to the same entity. **Modern Approaches** - **Dense Retrieval**: Encode mention context and entity descriptions into vectors, retrieve by similarity. - **LLM-Based**: Use large language models to disambiguate in-context. - **Autoregressive**: Models like **GENRE** generate entity names token by token conditioned on context. **Tools and Systems** - **spaCy** with entity linking components - **REL (Radboud Entity Linker)** - **BLINK** (Facebook/Meta) - **DBpedia Spotlight** Entity linking is a foundational building block for **knowledge-grounded AI** systems that need to reason about real-world entities.