Relation extraction is the NLP task of identifying and classifying semantic relationships between entities mentioned in text. Given a sentence like "TSMC manufactures chips for Apple," relation extraction would identify the manufactures_for relationship between the entities TSMC and Apple.

How Relation Extraction Works

• Input: Text containing two or more identified entities (from named entity recognition).
• Output: The type of relationship between entity pairs, selected from a predefined set (e.g., works_at, located_in, manufactures, acquired_by).
• Example: "Jensen Huang founded NVIDIA in 1993" → (Jensen Huang, founded, NVIDIA)

Approaches

• Supervised Classification: Train a model (BERT + classification head) on labeled examples of entity pairs and their relations. High accuracy but requires extensive annotated data.
• Distant Supervision: Automatically generate training data by aligning a knowledge base (like Wikidata) with text. If (TSMC, headquartered_in, Hsinchu) is a known fact, any sentence mentioning both "TSMC" and "Hsinchu" is assumed to express that relation.
• Few-Shot / Zero-Shot: Use LLMs to extract relations with minimal or no training examples by providing instructions and demonstrations in the prompt.
• Open Relation Extraction: Extract relation phrases directly from text without constraining to a predefined schema (see open information extraction).

Challenges

• Ambiguity: The same entity pair can have multiple relations depending on context.
• Long-Range Dependencies: Relations may span multiple sentences or require coreference resolution.
• Domain Adaptation: Models trained on general text may not handle domain-specific relations (semiconductor manufacturing, legal contracts) without adaptation.

Applications

Relation extraction is essential for knowledge graph construction, question answering, document understanding, and intelligence analysis. It transforms unstructured text into structured knowledge that can be queried, reasoned over, and integrated into AI systems.