We thank the authors of PPTOD for gathering up the publicly available human-written multi-turn dialogue corpora.

Preparation of Benchmark TOD Task Datasets:

MultiWOZ Data:

The MultiWOZ dataset is used for both end-to-end task-oriented dialogue modelling and dialogue state tracking tasks.

(1) Preparation:

To acquire the processed dataset, you can run the following commands.

cd ./multiwoz/
chmod +x ./download_data.sh 
chmod +x ./data_preparation2.1.sh 
chmod +x ./data_preparation2.2.sh 

sh ./download_data.sh # download the necessary data
sh ./data_preparation2.1.sh # prepare the data for multiwoz 2.1
sh ./data_preparation2.2.sh # prepare the data for multiwoz 2.2

Take a coffee, this process will take around 60 minutes.

(2) Data Format:

[
    {
        "dial_id": "PMUL1170",
        "user": "i need to take a train out of cambridge , i will be leaving town on wednesday .",
        "resp": "there are [value_choice] trains out of [value_departure] on [value_day] . do you have a departure time in mind ?",
        "bspn": "[train] day wednesday departure cambridge",
        "aspn": "[train] [inform] choice departure day [request] leave",
        "turn_num": 0,
        "db": "[db_3]",
    },
    {
        "dial_id": "PMUL1170",
        "user": " i would like to go to peterborough and leave after 12:45 , i have to attend a meeting beforehand .",
        "resp": "[value_id] leaves at [value_leave] on [value_day] . will that work for you ?",
        "bspn": "[train] day wednesday departure cambridge leave 12:45 destination peterborough",
        "aspn": "[train] [inform] day leave id",
        "turn_num": 1,
        "db": "[db_3]",
    },
    ...
]

We use json to store the data. Each dialogue session is represented as a list of turns. Each turn is represented as a dictionary that contains the following fields:

• dial_id - The unique ID for the dialogue session instance.
• user - The user's utterance.
• resp - The delexicalized reference system response.
• bspn - The belief state.
• aspn - The system action.
• turn_num - This argument indicates the turn position in the dialogue session, e.g., if turn_num = 0 means this is the very first turn in the whole dialogue session.
• db - The database query result.

Pre-training Corpora Preparation:

We download the raw data of several publicly available human-written multi-turn dialogue corpora collected by ToD-BERT. To run the pre-training data preparation scripts, please first install gdown library as:

pip3 install gdown

Then, run the following commands to download the raw data.

cd pre-training_corpora
sh download_raw_data.sh

For the following processing of the raw data, please refer to Prompt-based Fine-tuning seciton in the main README.