📄 EMNLP 2026 · Open Source · 344 Tests

Turn Any Chinese Novel into an
Interactive Knowledge Graph

AI Reader extracts characters, relationships, locations, and events from novels spanning 100–2,500 chapters, then generates relationship graphs, procedural world maps, timelines, and encyclopedias — automatically.

Try Live Demo → Download Desktop App ⭐ GitHub

91.3%

Entity Precision

+28pp

Location Hierarchy Improvement

2,905

Chapters Analyzed

344

Automated Tests

Four Visualization Dimensions

From raw text to multi-dimensional interactive analysis — no manual annotation required.

📊

Relationship Graph

Force-directed graph with category-colored edges, path finding, density optimization for 400+ character novels.

🗺️

Procedural World Map

Constraint-solved layout with distance-field terrain, hand-drawn coastlines (roughjs), animated character trajectories.

📅

Event Timeline

Chapter-by-chapter event tracking with character swimlanes, importance filtering, and emotional tone linking.

📖

Character Encyclopedia

Cross-chapter profiles with relationship chains, alias resolution (行者/大圣/齐天大圣 → Sun Wukong), and scene index.

Six-Layer Neurosymbolic Validation

LLMs extract knowledge but make systematic errors. Our pipeline corrects them — 52% of spatial containment relationships had incorrect direction before validation.

Morphological Filtering

19 rules based on Chinese place-name structure (专名+通名)

Suffix-Rank Direction Fix

110+ geographic suffixes → hierarchical rank (洲 > 国 > 城 > 村 > 殿)

Multi-Source Voting

5 vote sources with temporal decay (later chapters weighted higher)

Transitivity Verification

O(V+E) parent-chain rank consistency check

Hierarchy Consolidation

12-step post-processing: synonym merge, orphan rescue, oscillation damping

LLM Reflection

Suspicious pairs sent to LLM for verdict with evidence requirement

Research

📄

Paper

EMNLP 2026 Demo Track

"Taming LLM Hallucinations in Long-Form Narrative Knowledge Extraction through Multi-Layer Validation"

Lei Feng, 2026

📊

ChiNovelKE Benchmark

First Chinese novel KE benchmark

480 human-annotated entries across 3 novels × 5 dimensions

View on Hugging Face →

📝

Citation

@inproceedings{feng2026aireader,
  title={AI Reader: Taming LLM
    Hallucinations in Long-Form
    Narrative Knowledge Extraction},
  author={Feng, Lei},
  booktitle={EMNLP 2026 Demos},
  year={2026}
}