Predicting word-level reading time in Romanian text by combining language models with the cognitive mechanics of reading.
Real data · participant 028 reading Solaris
Descoperirea lui Solaris avusese loc aproape cu o sută de ani înaintea nașterii mele. Planeta gravitează în jurul a doi sori: unul roșu, altul albastru. Timp de peste patru decenii, de ea nu s-a apropiat nici o navă.
- 39.19
- Final score / 100
- 36
- Baseline
The problem
People skip words, revisit earlier text, and pause to process information. Predicting reading time means accounting for both language context and individual reading behavior.
What we built
A two-stage hurdle model first predicts whether a word is skipped, then estimates reading time for words that are read. Lexical features, contextual RoBERT embeddings, and surprisal help model that behavior.
Under the hood
Character-offset tracking preserves alignment across transformer chunks. Lag and lead features capture spillover, while clause positions model sentence wrap-up effects.
