About the Embeddings Over Time Explorer

What this project is

The explorer lets you query models trained separately per year: find nearest neighbors for a word, plot semantic drift between word groups over time, project vectors onto interpretable axes, and run classic vector analogies. It is designed for researchers, students, and curious readers who want a hands-on view of distributional semantics over time—not just a static paper figure.

Data and training

Embeddings are trained with word2vec on year-conditioned slices derived from the FineWeb corpus. Articles are grouped into yearly subsets (including by year cues in URLs), then each year’s text is used to fit its own embedding space. The same interactive app runs on Hugging Face Spaces; models and vocabulary are under YearlyWord2Vec.

Web text reflects many languages and communities; the UI is in English, but the underlying slices are not limited to one dialect—more “babel du web” than a single textbook corpus.

What you can explore

• Nearest neighbors — cosine-similar words in a chosen year.
• Semantic evolution — how relationships between word groups change across years.
• Projection — 1D/2D views along semantic directions you define.
• Analogies — vector arithmetic (e.g. king − man + woman) with nearest-neighbor decoding.

Research context

The site supports a broader research program on quantifying language change with embedding spaces: comparing years, surfacing drift, and making the geometry of word2vec-style models tangible for teaching and replication.

Author and contact

Adam Eubanks
Brigham Young University
adameuba@byu.edu