poset_decoding

Poset Decoding

The official pytorch implementation of our paper Hierarchical Poset Decoding for Compositional Generalization in Language.

If you find our code useful, please consider citing our paper:

@inproceedings{Yinuo2020Hirearchical,
  title={Hierarchical Poset Decoding for Compositional Generalization in Language},
  author={Yinuo Guo and Zeqi Lin and Jian-Guang Lou and Dongmei Zhang},
  booktitle={Advances in Neural Information Processing Systems},
  year={2020}
}

Dependency

pip install -r requirements.txt

Data preprocess

get CFQ data : Download dataset from link

bash preprocess.sh

Training

Sketch Prediction

cd sketch_prediction/
bash ./train.sh

Traversal Path Prediction

The module is based on the open-source project Matchzoo-py https://github.com/NTMC-Community/MatchZoo-py

cd  ./traversal_path_prediction/MatchZoo-py/
python ./traversal_path_prediction/MatchZoo-py/train_esim.py

Evaluation

bash evaluate.sh

MCD2 and MCD3

In the aforementioned Training and Evaluation sections, we train and evaluate HPD on the MCD1 split.

To train and evaluate on MCD2/MCD3 split, please replace mcd1 to mcd2 or mcd3 in the following files:

• sketch_prediction/train.sh
• sketch_prediction/evaluate.sh
• traversal_path_prediction/MatchZoo-py/train_esim.py
• traversal_path_prediction/MatchZoo-py/evaluate_esim.py
• traversal_path_prediction/MatchZoo-py/datasets/cfq/load_data.py
• evaluate.sh

Acknowledgement

We will thank the following repos which are very helpful to us.

• Matchzoo-py

Contact

Any question please contact zeqi DOT lin AT microsoft DOT com