Narrative Knowledge Weaver: A Multi-Agent Framework for Knowledge Graph Construction and Analysis from Complex Narratives

Qiuyu Tian, Fengyi Chen, Youyong Kong, Fan Guo, Xin Zhang, Zhijing Xie, Yuyao Li, Jinjing Shen

Keywords: Knowledge, Graph, LLM Agent, Retrieval-Augmented Generation, Multi-Agent Systems

TL;DR: Narrative Knowledge Weaver: a multi-agent framework that builds narrative knowledge graphs with schema induction, disambiguation, and event-centric refinement, boosting QA over screenplays and novels.

Abstract:

Long-form narratives such as screenplays and novels require reasoning over evolving characters,
multi-stage events, and long-range temporal and causal structure.
Although recent LLM-based methods can extract surface entities and relations, automatically
induced knowledge graphs often lack the coherence and interpretability needed for narrative
understanding and downstream tasks such as continuity checking or character timeline analysis.
We introduce $\textbf{Narrative Knowledge Weaver}$, a multi-agent framework for constructing
high-quality, human-readable knowledge graphs from complex narratives.
The system combines adaptive schema induction, reflection-augmented extraction, and a
normalization-before-merge pipeline that performs type refinement, scope convergence, and
LLM-guided disambiguation.
A dedicated module conducts $\textbf{adaptive attribute enrichment}$ for narrative entities, aggregating
multi-granular evidence and reflection-guided feedback to incrementally refine and expand
schema-defined properties.
An $\textbf{event-centric refinement}$ stage further transforms raw event mentions into structured event
cards and causally organized Event Plot Graphs (EPGs).
All outputs are stored with fine-grained provenance and leveraged by a $\textbf{tool-augmented reasoning
agent}$ for temporal, causal, and structural queries.
Evaluations on Re-DocRED, a NarrativeQA-derived benchmark, and a Practitioner Screenplay QA
dataset show substantial improvements in entity normalization, relation accuracy, and event-level
reasoning over strong baselines including EDC, Hybrid Retrieval, and GraphRAG.
Beyond quantitative gains, the resulting graphs provide interpretable, application-ready
representations of story worlds, supporting detailed analyses of narrative dynamics—from
character states to causal chains and scene-level progression.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Submission Number: 9250