Author implementation of the following ACL 2021 paper.

1. Install an Anaconda virtual environment:

   conda create -n compgen python=3.7 anaconda
   

2. Install the following pip version

     pip install pip==20.0.2
   

3. Install the required dependencies:

   pip install -r requirements.txt
   

4. Install swi-prolog with version <8 (this is a Prolog interpreter used to execute GeoQuery). For example, for Ubuntu follow these instructions.
5. Download the datasets and extract the content to span-based-sp/datasets

The main script for running experiments is span-based-sp/run_span_exps.sh. To run an experiment use:

bash run_span_exps.sh domain

Where domain=[geo, scan, clevr]. A single experiment will train over all splits experimented in the paper for the specific domain. We use early stopping w.r.t denotation accuracy on the dev set. Results are written to a log file under the logs/ folder.

To conveniently print all logs, run:

python utils/combine_logs.py