Span-based Hierarchical Semantic Parsing for Task-Oriented Dialog

This is the codebase for the paper

Panupong Pasupat, Sonal Gupta, Karishma Mandyam, Rushin Shah, Mike Lewis, Luke Zettlemoyer.
Span-based Hierarchical Semantic Parsing for Task-Oriented Dialog.
EMNLP, 2019

Dependencies

• Python 3.x
• PyTorch 1.x
• PyYAML
• tqdm
• For the GloVe + LSTM model, download the processed GloVe vectors and extract the files to data/glove/.
• For the BERT model, 🤗Transformers is also required.

Basic Usage

• Training with the basic model (no edge scores) on debug data:

  ./main.py train configs/base.yml configs/data/artificial-chain.yml \
    configs/model/embedder-lstm.yml configs/model/span-node.yml

  This will train the model on the training data and evaluate on the development data. The results will be saved to out/___.exec/ where ___ is some number. The results contain the final config file, saved models, and predictions.

• The out/__.exec/__.meta metadata file stores the best accuracy and the epoch for that accuracy. To dump the metadata, use ./dump-meta.py out/__.exec/__.meta.

• To dump the predictions of a trained model in, say, out/42.exec/14.model on the test data:

  ./main.py test out/42.exec/config.json -l out/42.exec/14

Configs

• To use the real training data (TOP dataset), change the data config to configs/data/top.yml.

• To use edge scores, change the model config to configs/model/span-edge.yml.

• To use the BERT model, change both the data config and embedder config to the BERT counterparts (artificial-chain.yml > artificial-chain-bert.yml, top.yml > top-bert.yml, embedder-lstm.yml > embedder-bert.yml)

• The training accuracy will be incorrect because the decoder is not run during training. To turn on decoding during training, add

  -c "{"model": {"decoder": {"punt_on_training": false}}}"`

  to the command line.