UMAP of SBA coverage, 1940–1999

description

Articles mentioning the Swiss Bankers Association (SBA), mapped by semantic similarity and coloured by year, with clusters labelled by LLM. Every dot is an article; its colour is the year; clusters indicate semantic proximity. Zoom in and hover over articles to explore the map.

methodology

Corpus comprises all articles mentioning the SBA in Neue Zürcher Zeitung, Der Bund, Journal de Genève, and Gazette de Lausanne, published between 1940–1999.

Each article is first turned into vectors (256 dimensions) using the m-GTE embedding model. It is then projected to two dimensions using UMAP (cosine distance, normalised vectors); the coordinates on X and Y represent those two dimensions. Learn more about embeddings and UMAPs in this paper.

Clusters are identified by running HDBSCAN on the UMAP coordinates. An LLM (glm-5.3-flash) is then used to name each cluster from its 30 most representative articles.

The labels were manually checked against the articles of each cluster.

Check out the full pipeline here.

open full page back to explore