## TL;DR
Information Extraction (IE) transforms unstructured text into structured knowledge. Named Entity Recognition identifies people, organizations, and locations; Relation Extraction discovers connections between them. Modern LLMs have fundamentally changed IE — from specialized models to unified generative approaches.

## Core Explanation
IE pipeline: (1) Named Entity Recognition (NER) — locate and classify named entities (Person, Organization, Location, Date); (2) Relation Extraction (RE) — identify semantic relationships between entities (works_at, located_in, founded_by); (3) Event Extraction — detect event triggers and their arguments; (4) Coreference Resolution — link pronouns to entities. Traditional approach uses LSTM-CRF or BERT-based taggers; modern LLM approach uses instruction-following or code-generation formats.

## Detailed Analysis
Generative IE: prompt LLM to output JSON {"entities": [...], "relations": [...]}. Advantages: handle nested/overlapping entities naturally, zero-shot transfer to new entity types, unified architecture. Multimodal IE extends to documents with tables, forms, and images via LayoutLMv3, Donut, and Nougat. OCR+LLM hybrid pipelines (2025) combine traditional OCR with LLM correction. Applications: scientific literature mining, legal document analysis, financial report extraction.

## Further Reading
- SpaCy NER and Transformers
- HuggingFace Token Classification
- Awesome-LLM4IE Papers GitHub