SlotRefine: A Fast Non-Autoregressive Model for Joint Intent Detection and Slot Filling

Modified by: Steeve Laquitaine

Original reference:

Main paper to be cited (Di Wu et al., 2020)

@article{wu2020slotrefine,
  title={Slotrefine: A fast non-autoregressive model for joint intent detection and slot filling},
  author={Wu, Di and Ding, Liang and Lu, Fan and Xie, Jian},
  booktitle={EMNLP},
  year={2020}
}

Prerequisites

• THUMT codebase
• Conda

Original setup and train

• Original setup:

setup.sh      # create virtual env. and install dependencies

Train on atis dataset:

train.atis.sh # train model

Train on snips dataset:

train.snips.sh

Quick runs for testing

1. Configure parameters.yml and catalog.yml in conf/
2. Run a pipeline:

• train:

# python models.py --pipeline train

• predict:

# python models.py --pipeline predict

• Shuffle corpus:

python thumt/scripts/shuffle_corpus.py --corpus "data/atis/train/data" --seed 0 --num_shards 1

..and other scripts:

• build_vocab
• checkpoint_averaging
• convert_old_model
• convert_vocab
• input_converter
• shuffle_corpus
• visualize

Specs and stats

• Specs original paper:

  • decoding:
    • single Tesla P40 GPU

• Stats:

  • inference latency: 3.02 ms

• Stats:

  • train: 4 hours (200 epochs)

TODO

• complete variabilisation of configurat in utils.get_params()