README.md

metadata

task_categories:
  - text-generation
language:
  - en
size_categories:
  - 10K<n<100K

This dataset is based on the "cumulative" configuration (i.e., slots previously filled are kept in the state) of the MultiWoz 2.2 dataset available also at pfb30/multi_woz_v22. Therefore, the system and user utterances, the active intents, and the services are exactly the same. In addition to the data present in version 2.2, this dataset contains the annotations from versions 2.1, 2.3, and 2.4 for each dialogue turn. This dataset is an artefact of the paper Diable: Efficient Dialogue State Tracking as Operations on Tables (Lesci et al., 2023).

Data Format

A brief recap on the structure of the MultiWoz dataset:

• Each dialogue turn is composed of a system utterance and a user utterance, in this exact order.

• The initial system utterance is filled in with the none string.

• In the last dialogue turn, it is always the system that greets the user. This last turn is kept, and the user utterance is filled in with the none string. Usually, during evaluation, this dialogue turn is not considered.

• To save data as an arrow file, you must "pad" the states so they all have the same keys. To do this the None value is introduced. Therefore, when you load it back, it is convenient to have a way to remove the padding. To do so, a function like the following can help

def remove_empty_slots(state: dict[str, list[str] | None] | None]) -> dict[str, list[str]] | None]:
  if state is None:
      return None
  return {k: v for k, v in state.items() if v is not None}

• The schema has been updated to make all the versions compatible. Basically, the "book" string has been removed from slots in v2.2. The updated schema is the following

attraction-area
attraction-name
attraction-type
hotel-area
hotel-day
hotel-internet
hotel-name
hotel-parking
hotel-people
hotel-pricerange
hotel-stars
hotel-stay
hotel-type
restaurant-area
restaurant-day
restaurant-food
restaurant-name
restaurant-people
restaurant-pricerange
restaurant-time
taxi-arriveby
taxi-departure
taxi-destination
taxi-leaveat
train-arriveby
train-day
train-departure
train-destination
train-leaveat
train-people