The Repository for the EMNLP 2025 Paper The Medium is Not the Message: Deconfounding Document Embeddings via Linear Concept Erasure

The current pipeline generates and saves plots for the following visualizations:

• K-means clusterings, both before and after LEACE erasure;
• PCA projections, before and after LEACE erasure;
• Top-k retrieval results, before and after erasure;
• Total # of exact pairs, before and after erasure.

In addition, we provide all data used in our paper.