Final Project: 6.884 Neurosymbolic models for NLP

This repository contains our work for our project "We Put the Symbolic in Neurosymbolic: Tackling Compositionality by Learning Rules From Data". The work can be broadly divided into three parts:

1. Automatically learning intermediate representations using DreamCoder.
2. Testing the benefits and theoretical limitations of intermediate representations via the InterSCAN variant to SCAN.
3. End-to-end symbolic learning via program synthesis.

Our experiments testing on SCAN and InterSCAN can be found in seq2seq.py, mcd.py, utils.py, and launch.sh. The different splits can be found in the splits directory. For example, splits/around_right/train_inter.txt contains the training set for the around right split using intermediate representations. Code for our experiments automatically learning intermediate representations and the end-to-end program induction approach can be found in scan_dreamcoder/. Note: Experiments were conducted with a different fork of ec, so files are not fully integrated with the DreamCoder code base as presented here.

Output files from using dreamcoder are found in scan_dc_out. Output files from training on SCAN and InterSCAN are in out. Saved models and checkpoints are in saved.