Author Attribution (AA) Explainability Tool

This demo helps you see inside a deep AA model’s latent style space.

Currently you are inspecting LUAR with pre-defined AA tasks from the HRS dataset

🔍

Visualize

Place your AA task with respect to other background authors.

✏️

Generate

Describe your investigated authors' writing style via human-readable LLM-generated style features.

⚖️

Compare

Contrast with Gram2Vec stylometric features.

Select a model and a task source (pre-defined or custom)
Click Load Task & Generate Embeddings to load the task and generate embeddings
Run Visualization to see the mystery author and candidates in the AA model's latent space
Zoom into the visualization to select a cluster of background authors
Pick an LLM feature to highlight in yellow
Pick a Gram2Vec feature to highlight in blue
Click Show Combined Spans to compare side-by-side

Choose a Model to inspect

gabrielloiseau/LUAR-MUD-sentence-transformers gabrielloiseau/LUAR-CRUD-sentence-transformers miladalsh/light-luar AnnaWegmann/Style-Embedding Other

Select Task Source

Predefined HRS Task Upload Your Own Task

Pick a pre-defined task to investigate (a mystery text and its three candidate authors)

Choose which mystery document to explain

Click the button below to load the tasks and generate embeddings using selected model.

Run visualization to see which author is similar to the mystery document.

Visualization

What am I looking at?

This plot shows the mystery author (★) and three candidate authors (◆) in the AA model’s latent space.
The grey ● symbols represent the background corpus—real authors with diverse writing styles. You can zoom in on any region of the plot. The system will analyze the visible authors in that area and list the most representative writing style features for the zoomed-in region.
Use this to compare your mystery text’s position against nearby writing styles and investigate which features distinguish it from others.

Zoom in on the plot to select a set of background authors and see the presence of the top features from this set in candidate and mystery authors.

Quick Region Selection

Select a precomputed region to analyze, or zoom manually on the plot above

Precomputed Regions

Select a region to automatically zoom and analyze

None

LLM-derived style features prominent in the zoomed-in region

LLM-derived style features for this zoomed-in region

Gram2Vec Features prominent in the zoomed-in region

Features shown with normalized z-scores

Gram2Vec features for this zoomed-in region

Click "Show Combined Spans" to highlight the LLM (yellow) & Gram2Vec (blue) feature spans in the texts

LLM feature

Gram2Vec feature