Narrative entity extraction

Definition

Narrative entity extraction (NER applied to fiction) identifies named and structural elements in the manuscript: characters (including nicknames and variants), places, organizations, significant objects, datable events, sometimes relational states ("X is Y's brother"). It is the first step in any contextual intelligence pipeline: without reliable entities, neither a Narrative knowledge graph, structured AI contextual memory, nor Automated consistency analysis can produce usable results.

Approaches range from classical NER (models trained on general corpora — fast but imprecise on fiction) to LLM extraction (richer, costlier, hallucination-prone) to schema-guided extraction aligned with the author's writing bible. The ideal combines automatic extraction with human validation: the author fixes homonyms, merges duplicates, and confirms relations before they feed the graph.

What it's for

This is the foundational brick of a living codex. Extraction turns 400 pages of prose into queryable data: "how often does this character appear?", "was this place already named with a different spelling?", "was this object introduced before it matters at the climax?" It frees the author from exhaustive manual rereading for factual checks.

For the Narrative copilot (AI), extraction ensures suggestions rest on real entities from the text, not invented names. For Internal coherence, it supplies anchor points (ages, positions, possessions) that consistency rules can then compare. Re-running extraction after major rewrites is cheaper than manually rebuilding a bible from scratch — provided the author reviews edge cases the model mishandles.

Variants by genre

• Fantasy / SF: rich onomastic extraction — invented names, titles, species, artifacts; high risk of confusion with everyday vocabulary.
• Mystery: event-oriented extraction — murders, movements, clue discoveries; each fact tied to a moment and witnesses.
• Romance: relational extraction — character pairs, relationship states implied in dialogue and action beats.
• Literary fiction: selective extraction — core characters and symbolic places; avoid extracting every street mentioned in passing.
• Series: cross-volume extraction with disambiguation — the "John" in book 3 is not book 1's John without an explicit link.

Common pitfalls

• False positives: real place names confused with first names, cited brands treated as story entities.
• No disambiguation: three characters named "Marie" merged into one node.
• Schema too broad: noise extraction (every action verb becomes an "event").
• Blind trust in LLM: invented relations polluting the Narrative knowledge graph.
• One-shot extraction: snapshot at chapter 10 while chapters 1–9 were rewritten — stale data.
• Ignoring variants: "Mr. Dupont," "Henri," and "the old man" not linked to the same character.
• Treating extraction as a one-time import: fiction evolves scene by scene; entity lists need incremental refresh like any other project artifact.

Examples

• Harry Potter — hundreds of side characters and locations; structured extraction would map recurring appearances (e.g., Peeves, Firenze) to avoid cross-volume omissions.
• The Name of the Rose — Latin names, abstract places (library), ecclesiastical hierarchy: an extraction schema adapted to the medieval setting is essential.
• Big Little Lies — relational entities (parent group, shared secrets) matter more than places; extraction must capture alliances and exclusions, not only proper nouns.

Writing tips

• Manually check the first 20 extractions: nicknames, aliases and titles generate duplicates ("John" / "Dad" / "Mr Smith"). Without a manual pass, your index is wrong.
• Configure a frequency threshold: an entity cited once isn't a character. Filter to minimum 3 occurrences to avoid graph pollution.
• Distinguish entity from role: "the king" may refer to 3 different characters across a saga. Never merge without scene context.
• Re-extract after every major revision: a cut chapter can remove the only mention of a supporting character → ghost detection.

Learn more: Narrative threads: stop forgetting your story's promises and Why the Relationship Graph is the Secret Weapon for Complex Novel Authors.

Revision checklist

• Do Narrative knowledge graph and this notion reinforce each other, or does one contradict the other?
• Could a beta reader name Narrative entity extraction after reading without being told?
• Am I keeping a scene out of attachment that doesn't illustrate this notion?
• Did my last pass fix structure before style (see Line editing / polishing if needed)?
• Does the climax or pivot scene clearly show what Narrative entity extraction changes in the story?