Language Models Use Monotonicity to Assess NPI Licensing

This repository contains the code for the experiments of the ACL 2021 Findings paper "Language Models Use Monotonicity to Assess NPI Licensing".

The experimental pipeline can be run with the following command: python3 npi_pipeline.py -c npi_pipeline.json

For training our language models we used the script from the repository of Gulordava et al. (2018). The 37 models we have trained for our experiments can be downloaded here (9.9GB zip).

Our pipeline makes use of the diagNNose library, that can be installed using pip.