From embeddings to labelled maps

description

From embeddings to labelled maps: embeds articles, projects them with UMAP, clusters them, names the clusters with an LLM and exports an interactive map.

methodology

Articles are embedded with the Impresso embedding service and reduced to two dimensions with UMAP.

HDBSCAN finds the clusters; an LLM labels each one from a sample of its articles.

The notebook then checks the labels (newspaper and decade shares per cluster, a skim of articles, stability under other clustering settings) and exports the interactive map, which shows only the first 15 words of each article.

open full page back to explore