Our paper

Dialogue State Tracking with Multi-Level Fusion of Predicted Dialogue States and Conversations

DST-DCPDS

We develop our code baed on CHAN.

Structure

Requirements

• python3.6

• Install python packages:

  pip install -r requirements_train.txt

Usages

Data preprocessing

We conduct experiments on the following datasets:

• MultiWOZ 2.0 Download to get MULTIWOZ2.zip
• MultiWOZ 2.1 Download to get MULTIWOZ2.1.zip

We use the same preprocessing steps for both datasets. For example, preprocessing MultiWOZ 2.1:

cd /home/user/DST-DCPDS
# download multiwoz 2.1 dataset
# preprocess datasets for training DST and STP jointly
$ unzip -j MULTIWOZ2.1.zip -d data/multiwoz2.1-update/original
$ cd data/multiwoz2.1-update/original
$ mv ontology.json ..
$ python convert_to_glue_format.py

Pretrained BERT Download

• Model Download to get bert-base-uncased.tar.gz
• Vocab Download to get bert-base-uncased-vocab.txt

  cd /home/user/DST-DCPDS
  # rename model file
  mv bert-base-uncased.tar.gz bert-base-uncased.model
  # mv file to target file
  mv bert-base-uncased.model /home/user/DST-DCPDS/data/pytorch_bert/.pytorch_pretrained_bert/.
  mv bert-base-uncased-vocab.txt /home/user/DST-DCPDS/data/pytorch_bert/.pytorch_pretrained_bert/.

Train

Take DST-DCPDS for both-level training as an example:

• Train

  bash run_multiwoz2.1_all.sh

For teacher-forcing set the mix_teaching_force to 0

For uniform scheduled sampling set the mix_teaching_force to 1

Evaluation

Take DST-DCPDS for both-level training as an example:

General performance

• uncomment the last line in the run_multiwoz2.1_all.sh

• run sh

  bash run_multiwoz2.1_all.sh

  Related-slot

• select taxi domain dialogues from the test datasets to generate file select_taxi.tsv

  python select_taxi_dialouge.py

• move selected dialogue to target folder

  mv select_taxi.tsv data/multiwoz2.1-update/.

• Manually replace the explicit expression to implicit form, for example replace rice house in the utterance to restaurant. You can also skip the steps.

• replace the evaluation file name in code\main.py from test.tsv to select_taxi.tsv

• run the prediction process

  bash run_multiwoz2.1_all.sh

• run evaluation process

  python eval_related_slot.py

  Value-delete

• select dialogues, which has value-delete phenomenon, from test datasets to generate file selected_dialog_cand.json

  python select_value_delete.py

• Manually replace the slot value in the dialogue, which has value-delete phenomenon with [slot-name], for example replace rice house with [resturant-name]. The template file name as augment_dialog_cand.json

• Augment dialogue to generate file update_test.tsv

  python augment_value_delete.py

• replace the evaluation file name in code\main.py from test.tsv to update_test.tsv

• run the prediction process

  bash run_multiwoz2.1_all.sh

• run evaluation process

  python eval_value_delete.py