DiscourseGeometry

Compare two collections of texts on theoretically defined semantic axes, using the spatial-geometric framework of Vallego (2026): centroid projection (where does each collection sit, and did it move?), Frobenius norm (how large is the difference, and how much of it do the named axes capture?) and eigenanalysis (did the shape and orientation of the collection change?).

Each axis is defined by two sets of exemplar texts (pole A and pole B). The first axis is the primary axis; further secondary axes (up to five) decompose it and are orthogonalised in the order given. Use Paired for two sets of responses to the same prompts in the same order (e.g. a base model and its fine-tuned version), and Independent for two corpora. Separate units with a line containing only ---, with blank lines, or one per line.

Part of the H4rmony Project · Cite this work (expand below)

Preset

Loads axes, exemplars and example collections

Units are separated by

Primary axis: name

Pole A label (positive)

Pole B label (negative)

Pole A exemplars

Pole B exemplars

Secondary axis 1: name

Pole A label (positive)

Pole B label (negative)

Pole A exemplars

Pole B exemplars

Secondary axis 2: name

Pole A label (positive)

Pole B label (negative)

Pole A exemplars

Pole B exemplars

Secondary axis 3: name

Pole A label (positive)

Pole B label (negative)

Pole A exemplars

Pole B exemplars

Primary axis

Secondary axes

3D view

Semantic map: the chosen x axis on its 0–100% scale, with its pole exemplars; y and z are the main directions in which all texts differ apart from that axis (as in SpatialSemanticAnalysis). Chosen axes: x, y and z are three axes of your choice.

Semantic map: x = one axis, y and z = main remaining dimensions Chosen axes: any three axes

x axis

y axis (chosen axes only)

z axis (chosen axes only)

Orthogonalise in the order x → y → z

Per-unit shift

Units (first 500; full table in the CSV)

Downloads: per-unit CSV and all results (JSON)