Analogical Matching Network

A neural-based system for outputting analogical matchings. It combines recent advances in Graph Neural Networks, Transformers, and Pointer Networks.

Experiment Setup

To set up experiments

sh setup_directories.sh
pip3 install -r requirements.txt (ensures package requirements are met)

There is a pretrained model included in the models directory, thus training is not necessary

Prior to Training

Before training, you should generate the synthetic dataset

  python3 generate_dataset.py

This also builds the graph encoder object that goes along with the synthetic dataset, which is needed for training

Training a Model

To train a model

  python3 train_model.py

all results and intermediate progress is going to be stored in ./results, so you should reference the files in there to see how your model is progressing as it trains. All models are stored in ./models, with latest_model_obj.pkl being the model called in the testing and visualization code

The default settings are set to those listed in the paper.

Testing a Model

To test a model

  python3 test_model.py

test results are stored in ./results

Visualizing a Model

To visualize outputs

  python3 visualize.py --domain <domain_name>

where <domain_name> is the domain you want to visualize analogies for, e.g. --domain atom. The 'custom' domain allows you to enter your own base and target logical expression sets. An example of its use is

 python3 visualize.py --domain custom

 Please enter expressions for the Base at the prompt...
 Type END to finish...
 >>> (p (f a b))
 >>> (p (f c d))
 >>> END

 Please enter expressions for the Target at the prompt...
 Type END to finish...
 >>> (p (f b a))
 >>> (p (f d c))
 >>> END

Citing this Work

If you end up using this work in a publication, please cite the original paper

  @article{crouse2020neural,

     title={Neural Analogical Matching},

     author={Crouse, Maxwell and Nakos, Constantine and Abdelaziz, Ibrahim and Forbus, Kenneth},

     journal={arXiv preprint arXiv:2004.03573},

     year={2020}

  }

hyzcn / NeuralAnalogy