This repository contains the code used in the paper "Representation of linguistic form and function in recurrent neural networks", the version published in Computational Linguistics.

The structure is the following:

• doc: Latex sources and figures for the paper
• models: RNN models trained on MS COCO (need to be downloaded from https://doi.org/10.5281/zenodo.825407)
• data: data files
• src: code for analysis
  • src/omission.py Omission scores
  • src/depparse.py Dependency parse using Spacy
  • src/mutual.py Mutual information scores
  • src/example.py Examples
  • src/ridge.py Ridge regression models
  • src/analysis.R R code genarating figures
  • src/analysis_mi.R R code genarating figures

The models were trained using the following software: https://github.com/gchrupala/reimaginet/releases/tag/v0.4