Solving Arithmetic Word Problems by Scoring Equations with Recursive Neural Networks

Introduction

This repository contains the code to reproduce the results from the following paper:

@article{zaporojets2021solving,
  title={Solving arithmetic word problems by scoring equations with recursive neural networks},
  author={Zaporojets, Klim and Bekoulis, Giannis and Deleu, Johannes and Demeester, Thomas and Develder, Chris},
  journal={Expert Systems with Applications},
  volume={174},
  pages={114704},
  year={2021},
  url={https://doi.org/10.1016/j.eswa.2021.114704}
  publisher={Elsevier}
}

Downloading the Dataset and Models

The dataset and models are located in this link. Once downloaded and unzipped, the directory structure should look as follows:

├── data/
│   ├── embeddings/
│   └── single_eq/
├── experiments/
│   ├── models/
│   ├── params/
│   └── results/
├── src/
...

Creating the Environment

It is recommended to create a separate conda environment, then install the necessary packages inside:

conda create -n mwp python=3.7.10
conda activate mwp 
pip install -r requirements.txt
python3 -m spacy download en_core_web_sm

Reproducing the Results

To reproduce the results of the paper (Tables 4-5 and 7), run the following script:

python src/main_reproduce_results.py

By default, the script will read the predictions in .csv files located in experiments/results/ directory. To execute all the models, pass --execute_models True parameter to the script. To execute the models on GPU, pass --use_gpu True parameter.

Training the Models

Work in Progress

Contact

You can contact us at klim.zaporojets@ugent.be