gpalrepo / Natural_Language_Deduction_with_Incomplete_Information

Natural Language Deduction with Incomplete Information

Generating premises in underspecified settings through abduction.

https://arxiv.org/abs/2211.00614

Overview

This repository holds the code for the paper Natural Language Deduction with Incomplete Information

All the data for training and evaluating models can be found in {PROJECT_ROOT}/data

The newly introduced Everyday Norms: Why Not? dataset can be found in {PROJECT_ROOT}/data/full/enwn/enwn.json (to visualize it check the README in src/visualization_tools)

For examples on how to call our Step Models, Heuristics, etc. check {PROJECT_ROOT}/examples.ipynb for a brief overview.

Otherwise, all the major sections of our system found in {PROJECT_ROOT}/src are either heavily documented or have a README inside their root folder meant to help guide the reader towards using our code.

Data

The entailmentbank and Everday Norms: Why Not? datasets can be found in data/full

Installation

1. virtualenv venv we have tested with python 3.8
2. source venv/bin/activate
3. pip install -r requirements.txt
4. python ./install.py (expect this to take a long time based on your internet speed)

Step 4 will download all the models needed to recreate the experiments from the paper as well as two small step models that also perform well. (We use the smaller step models for testing/debugging and the 3billion parameter models for experiments.)

You can see all the models that can be downloaded via python ./install.py -s

And you can install them individually via python ./install.py -a abductive_heuristic abductive_step_model_small ...etc

Getting started (recreate paper results)

1. Follow the examples.ipynb
2. Go to src/experiments and follow the README
3. to recreate the paper results make sure to update the config folder with the right parameters.

Folder breakdown

• configs: parameters used for experiments (templates for the configs of the experiments we used in our paper)
• data: All training and evaluation data we used in the paper
• output: where the experiments are set to put the data/output/results of a run (check src/experiments)
• src: All source code for our project broken up into submodules.
• trained_models: Important folder where all trained models should be stored (put any checkpoints downloaded into this folder)

Requirements

All experiments were performed on the following specs (some may not matter)

• Python Version: 3.8.3
• tested on Mac and Linux
• Transformers version 4.20.0 (really important)

Make sure you run pip install -r requirements.txt

Authors

• Zayne Sprague
• Kaj Bostrom
• Swarat Chaudhuri
• Greg Durrett