neural_constructions

This repository contains code and data for "Investigating verb bias in neural language models", to appear in EMNLP.

Our analyses and predictions were pre-registered on the OSF.

• DATIVE contains our new dataset of 50K human judgements, as well as code to reproduce human data collection and all subsequent analyses.
• SWBD contains materials for our re-analysis of the natural corpus of datives annotated by Bresnan & Nikitina (2007)
• models is a module used by our other analyses to extract sentence probabilities and hidden states from all neural language models we consider