SynapTale Demonstrates Temporal Knowledge Graph for 180,000-Word Oz Series

Temporal Knowledge Graph for Narrative Analysis

SynapTale has unveiled a demonstration of its temporal knowledge graph technology, applied to the first 100 chapters of the Oz series, totaling 180,000 words. This graph models the story as a network of nodes, representing entities, and edges, representing their actions and relationships. The system allows users to explore the story's state at every chapter, tracking how elements evolve over time.

System Capabilities and Data Points

The current demo includes 232 entities and 1,852 edges. A key feature is its ability to retain information over long narrative spans; for example, a promise made in chapter 8 is still recognized by chapter 100. The platform also provides analytics, revealing specific narrative details such as character kill counts, deception patterns, and persistent plot elements like the Scarecrow’s debt to a stork.

Technical Implementation

The underlying technology involves five distinct multi-agent pipelines combining Large Language Models (LLMs) and Natural Language Processing (NLP). These pipelines handle prescan, ontology construction, chapter-by-chapter graph extraction, retrospective validation, and linguistic analysis. The system employs a specific ontology scan to define story-specific entity and edge types, preventing LLMs from generating inconsistent or irrelevant fields.

To manage the temporal aspect of a narrative, the graph uses three types of edges: 'event' for instantaneous actions, 'identity' for facts, and 'state' for persistent actions that require justification for termination. This design ensures scalability by limiting continuously active edges to a minority of state and identity types.