Surprisal values from XGLM models

This repository contains the code to compute surprisal values from XGLM model. We used the MECO corpus and the XGLM model family to assess the relationship between the psychological accuracy of a language model (namely, the capability of a surprisal estimate to explain variance in human responses) and its linguistic accuracy (i.e., its ability to accurately predict the next token).

Useful resources

• :eyes: Eye-tracking data:
  • MECO-L1 corpus (paper|data).
• :computer: Computational models
  • XGLM model family (paper|model)