SLU-Aug-PrLM

The code for Interspeech2021 paper "Augmenting Slot Values and Contexts for Spoken Language Understanding with Pretrained Models"

Getting Started

1. Requirements

• Python 3.6
• Pytorch == 1.4.0
• transformers ==2.5.1
• tensorboard == 2.5.0

2. Resources

1. Download bart-large pretrained model and put it into resources/ folder.
2. Substitute the modeling_bart.py file in the transformer package with the given file utils/modeling_bart.py

3. Training

Augmenting Slot Values

Refer to the bash file utils/run_bart_single.sh (You need to enter the directory according to your settings in the bash file)

1. First, choose the training dataset size by

   size=327 # for snips you can choose {327, 1308, 13084}
   

2. Fine-tune the bart-large model with the given dataset.

   python run_delex_to_raw_single.py --train_data_file ${SNIPS_data}train_${size}_raw.txt --train_delex_file ${SNIPS_data}train_${size}_delex.txt --train_label_file ${SNIPS_data}train_${size}_label.txt --output ../result/snips/single_${size}/  --cache_dir cache/ --do_train --num_train_epochs 5 --logging_steps 1000 --save_steps 0 --overwrite_output_dir --overwrite_cache --seed ${seed} --cuda_id ${gpu_id} --model_name_or_path ../resources/bart_large
   

3. Run the fine-tuned model and generate augmented data.

   python get_delex_to_raw_single.py --gen_data_file ${SNIPS_data}train_${size}_raw.txt --gen_delex_file ${SNIPS_data}train_${size}_delex.txt --gen_label_file ${SNIPS_data}train_${size}_label.txt --output_file ../result/snips/single_${size}/checkpoint-${checkpoint}/gen_data.txt --output_delex_file ${SNIPS_data}train_${size}_single_delex.txt --output_label_file ${SNIPS_data}train_${size}_single_label.txt --model_type bart --model_name_or_path ../result/snips/single_${size}/checkpoint-${checkpoint}/ --seed ${seed} --cuda_id ${gpu_id}
   

4. Filter the unqualified data and save the augmented data in ../result/snips/single_${size}/checkpoint-​${checkpoint}/filtered_data.txt

   python filter_single_data.py --delex_file ${SNIPS_data}train_${size}_single_delex.txt --raw_file ${SNIPS_data}train_${size}_raw.txt --label_file ${SNIPS_data}train_${size}_single_label.txt --gen_file ../result/snips/single_${size}/checkpoint-${checkpoint}/gen_data.txt --filter_file ../result/snips/single_${size}/checkpoint-${checkpoint}/filtered_data.txt --org_file ${SNIPS_data}train_${size}_
   

5. Use LSTM or BERT to test the result of augmented data.

Augmenting Contexts

Refer to the bash file utils/run_bart_act.sh (You need to enter the directory according to your settings in the bash file)

The process is consistent with the one in Augmenting Slot Values

4. Others

We also provide the link of other baselines methods for comparison

• VAE
• Seq2seq

If you have any questions, please contact with [email protected]