Information Terra: A Narrative-Anchored Semantic-First Projection of Document Embeddings

Information Terra addresses the challenge of visualizing document corpora as coherent, navigable storylines. Unlike traditional dimensionality reduction techniques such as UMAP or PCA, which organize documents based on variance or neighborhood topology, Information Terra uses a semantic-first approach. By allowing an analyst to select two endpoint documents, the system defines a specific narrative arc. It then projects all documents onto a sphere where the great-circle geodesic between these endpoints serves as the prime meridian. This geometry ensures that latitude directly corresponds to narrative progress, while longitude captures thematic deviation from the primary storyline.

The system operates in two distinct stages: projection and visualization. In the projection stage, document embeddings are mapped to a bipolar coordinate system on the hypersphere. A modified version of the Maximum Capacity Path (MCP) algorithm is used to extract a narrative trail that is constrained to be monotone in geodesic progress, preventing the storyline from backtracking. The visualization stage then transforms this spherical scatter plot into an Earth-like map. It uses kernel density estimation (KDE) to identify document clusters, which are rendered as continents and islands, and employs a small language model to automatically generate thematic labels for these landmasses.

By anchoring the visualization in a user-defined narrative, Information Terra provides a more intuitive and structured way to explore large document collections. The globe metaphor offers familiar navigational affordances, such as rotation and antipodal reading, which help analysts maintain context while exploring thematic extremes. Because the projection is deterministic and computationally efficient, it supports interactive sensemaking, allowing users to rapidly refine their narrative focus by adjusting endpoint documents.

Geodesic

The shortest path between two points on the embedding hypersphere, which in this system defines the primary narrative axis.

SLERP

Spherical linear interpolation, the mathematical method used to calculate the great-circle arc between two document embeddings.

Bipolar Projection

A coordinate system where two chosen documents serve as the north and south poles, organizing the remaining corpus based on their relationship to these points.

Narrative Trail

A sequence of documents extracted from the corpus that forms a coherent storyline, here constrained to move monotonically toward the target endpoint.

Thematic Deviation

The longitudinal distance of a document from the primary narrative geodesic, representing how much a document diverges from the main storyline.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.