Embedding Models

More models for latent networks.

Latent networks measure meaning by the similarity between word vectors. SemNA ships with GloVe-100 built in — download any of these to use a different pretrained model, then import the file into the app.

Available models

Each is a single pretrained word-embedding file. Download it here (your browser handles the download), then import it in SemNA. Larger models capture more, but take longer to download and load.

GloVe

Wikipedia + Gigaword · 100d · 400K words

A compact, general-purpose model. Good default for most corpora.

✓ Built into the app

GloVe (300d)

Wikipedia + Gigaword · 300d · 400K words · ~470 MB

The higher-dimensional GloVe — richer similarities than the built-in 100d version, still a modest download.

Download ↓

Word2Vec

Google News · 300d · 3M words · ~3.6 GB

Trained on ~100B words of news. Very large vocabulary; similarities tend to run lower. Words not in the training data are dropped.

Download ↓

FastText

Wiki-news, subword-aware · 300d · ~2.5 GB

Uses character n-grams, so it can produce a vector for any word — even rare terms and misspellings. Similarities tend to run higher.

Download ↓

Only static word-embedding models work with SemNA's latent networks (Word2Vec, GloVe, fastText, or a model you train yourself). Contextual models like BERT don't apply. You can also train a custom model on your own corpus with the semna Python package.

How to use a downloaded model

Download a model above (it lands in your Downloads folder).
In SemNA, on the Configure Network step, choose Latent Meaning.
Click Import a model, and pick the file you downloaded.
It's added to your model list — selected and ready to build. It stays available every time you open SemNA.

The file is copied into SemNA's own storage on import, so you can delete the download afterward.