The code of AAAI2021 paper Few-Shot Learning for Multi-label Intent Detection.

The code framework is based on few-shot learning platform: MetaDialog.

python >= 3.6
pytorch >= 1.5.0
transformers >= 2.8.0
allennlp >= 0.8.2
tqdm >= 4.33.0

Down the pytorch bert model, or convert tensorflow param yourself as follow:

export BERT_BASE_DIR=/users4/ythou/Projects/Resources/bert-base-uncased/uncased_L-12_H-768_A-12/

pytorch_pretrained_bert convert_tf_checkpoint_to_pytorch
  $BERT_BASE_DIR/bert_model.ckpt
  $BERT_BASE_DIR/bert_config.json
  $BERT_BASE_DIR/pytorch_model.bin

Set BERT path in the ./utils/config.py

Get data at ./data/

Set test, train, dev data file path in ./scripts/

Full data is available by contacting me, or you can generate it by your self:

We provide a generation tool for converting normal data into few-shot/meta-episode style. See details at here

Execute the command line to run with scripts:

source ./scripts/run_b_stanford_1_main.sh [gpu_id]

We provide all scripts for experiment at ./scripts/, and you can also directly run with ./main.py.

bert based scripts:

• run_b_stanford_1_main.sh
• run_b_stanford_5_main.sh
• run_b_toursg_1_main.sh
• run_b_toursg_5_main.sh

electra based scripts:

• run_e_stanford_1_main.sh
• run_e_stanford_5_main.sh
• run_e_toursg_1_main.sh
• run_e_toursg_5_main.sh

• script: scripts/get_tag_data_from_training_dataset.py
• operation:
  • change the parameters called MODEL_DIR and DATA_DIR
• command: python scripts/get_tag_data_from_training_dataset.py