How much data does it take to learn a language pattern?

Compare neural models trained on 100, 1,000, and 10,000 examples. See when more data helps—and when it doesn't.

The task: turn a word into the form you need

Given a dictionary form and grammatical features, predict an inflected word: sing + past tense → sang. Explore English, German, French, and Italian, from regular endings to irregular stem changes.

The models: build words one edit at a time

These transition-based neural transducers learn to copy, delete, insert, or substitute characters. Each language has three separately trained models, using the same architecture with different amounts of supervision. Training uses imitation learning with an edit-distance expert.

The exploration: look beyond a correct answer

Start with test examples where model predictions differ. Compare their outputs, inspect how they build a word, and check whether the word or feature combination appeared in training. Then try your own input.

Can a small model learn a productive rule? Does more data resolve an irregular form? Can the largest model still get it wrong?

Based on Makarov & Clematide (2020) · CoNLL–SIGMORPHON 2017 Task 1. Details and data links in Documentation.