linc's Introduction

LINC: Logical Inference via Neurosymbolic Computation

Repository for the paper LINC: A neuro-symbolic approach for logical reasoning by combining language models with first-order logic provers by Theo X. Olausson*, Alex Gu*, Ben Lipkin*, Cedegao E. Zhang*, Armando Solar-Lezama, Joshua B. Tenenbaum, & Roger P. Levy, to be presented at EMNLP 2023.

Code is provided to reproduce all experiments and figures.

Setup

Requirements: Anaconda, Make, Prover9

make setup

Usage

To rerun our exact experiments:

nano SUBMIT.sh # cfg for own cluster and submit contents of $JOB env variable
make run

To run custom experiments within our framework:

accelerate launch runner.py **kwargs
# see `eval/args.py` for options.

To replicate figures and tables from provided outputs:

touch outputs/run.done # don't rerun analyses
make analyze # core tables and figures
# see `analysis/notebooks` for supplemental tables and figures

Acknowledgements

This evaluation framework structure is adapted from the BigCode project and EleutherAI whom we thank for their contributions to open source.

Citation

@inproceedings{OGLZ_LINC_2023,
	author={Theo X. Olausson* and Alex Gu* and Ben Lipkin* and Cedegao E. Zhang* and Armando Solar-Lezama and Joshua B. Tenenbaum and Roger P. Levy},
	title={LINC: A neuro-symbolic approach for logical reasoning by combining language models with first-order logic provers},
	year={2023},
	journal={Proceedings of the Conference on Empirical Methods in Natural Language Processing},
}

Recommend Projects