COURSE 15L1100% FREE
Verified 2026-08-14

Document-level information extraction

**Document-level information extraction (doc-IE)** pulls structured facts—entities, relations, events—from **whole documents**, not only single sentences. Zheng, Wang, and Huang (FuturED 2024) survey contemporary doc-IE, analyze errors o...

What it is

Document-level information extraction (doc-IE) pulls structured facts—entities, relations, events—from whole documents, not only single sentences. Zheng, Wang, and Huang (FuturED 2024) survey contemporary doc-IE, analyze errors of strong systems, and highlight lingering blockers. SLP3’s information extraction chapter (relations, events, time) plus coreference chapters are the conceptual base.

<!-- IMAGE: multi-paragraph doc → entity graph with cross-sentence links -->

HIGH PRIORITYDIAGRAM / GRAPH
◷ IN PRODUCTION

Visual Spec & Architecture Diagram

Document-level IE: multi-paragraph fake medical/business note on left; extracted entity nodes (Person, Org, Drug) and relation edges (works_at, prescribed) forming a graph on right. Cross-sentence dashed arrow labeled 'coreference'.

Educational Focus: Shows why doc-level IE ≠ sentence NER—relations span context.

Why it matters

Contracts, scientific papers, incident reports, and news stories spread facts across paragraphs. Sentence-level NER/RE misses “who did what to whom” when arguments are pages apart. Chatbots that “remember the vibe” are not a substitute for structured IE with audit trails.

How it works (plain)

  1. Detect entities (and keep coreference clusters: “she” = “Dr. Ng”).
  2. Extract relations/events that may cross sentence boundaries.
  3. Use discourse context / document encoders / graph methods as systems evolve.
  4. Evaluate with precision/recall on document-annotated sets.
  5. Analyze errors: coreference failures and weak reasoning show up repeatedly in the 2024 survey’s findings.

Everyday example

In a 12-page vendor contract, the liability cap appears in section 9 while the party nickname is introduced in section 2. Doc-IE must link them; sentence NER alone will not.

Try it

Take a one-page news story. Write three triples (subject, relation, object) that require crossing sentence boundaries. Note which hinge on pronouns.

Myths

⚠️ Myth: Long-context LLMs solved doc-IE.
✓ Reality: They help draft candidates; surveys still find coreference and reasoning failures—and you still need schemas + eval.
⚠️ Myth: Sentence IE + concatenation is enough.
✓ Reality: Cross-sentence arguments and relation transitivity create distinct failure modes.
⚠️ Myth: Perfect entity tags imply perfect relations.
✓ Reality: Relation errors persist after entity detection.

Sources