Quenya Word Generator
Generate words with the MLP and compare it with a bigram baseline. Both run in your browser.
0.5 · more predictable1.5 · more random
Applies to both models on the next batch.
MLP
Batch 001A 2D embedding for each character; the previous 16 tokens pass through a 128-unit tanh layer to predict the next character.
Architecture based on Bengio et al. (2003).
Bigram baseline
Batch 001A 34 × 34 count matrix from the 4,414-word training split. Each new character depends only on the previous character. View matrix
Grey: attested or generated in an earlier batch.