task-oriented-dialogues/SimJointGEN/README.md

Simulator Generated Dataset (sim-GEN)

This directory contains an expanded set of dialogues generated via dialogue self-play between a user simulator and a system agent, as follows:

• The dialogues collected using the M2M framework for the movie ticket booking task (sim-M) are used as a seed set to form a crowd-sourced corpus of natural language utterances for the user and the system agents.
• Subsequently, many more dialogue outlines are generated using self-play between the simulated user and system agent.
• The dialogue outlines are converted to natural language dialogues by replacing each dialogue act in the outline with an utterance sampled from the set of crowd-sourced utterances collected with M2M.

In this manner, we can generate an arbitrarily large number of dialogue outlines and convert them automatically to natural language dialogues without any additional crowd-sourcing step. Although the diversity of natural language in the dataset does not increase, the number of unique dialogue states present in the dataset will increase since a larger variety of dialogue outlines will be available in the expanded dataset.

This dataset was used for experiments reported in this paper. Please cite the paper if you use or discuss sim-GEN in your work:

@article{liu2018dialogue,
  title={Dialogue Learning with Human Teaching and Feedback in End-to-End Trainable Task-Oriented Dialogue Systems},
  author={Liu, Bing and Tur, Gokhan and Hakkani-Tur, Dilek and Shah, Pararth and Heck, Larry},
  journal={NAACL},
  year={2018}
}

Data format

The data splits are made available as a .zip file containing dialogues in JSON format. Each dialogue object contains the following fields:

• dialogue_id - string unique identifier for each dialogue.
• turns - list of turn objects:
  • system_acts - list of system dialogue acts for this system turn:
    • name - string system act name
    • slot_values - optional dictionary mapping slot names to values
  • system_utterance - string natural language utterance corresponding to the system acts for this turn
  • user_utterance - string natural language user utterance following the system utterance in this turn
  • dialogue_state - dictionary ground truth slot-value mapping after the user utterance
  • database_state - database results based on current dialogue state:
    • scores - list of scores, between 0.0 and 1.0, of top 5 database results. 1.0 means matches all constraints and 0.0 means no match
    • has_more_results - boolean whether backend has more matching results
    • has_no_results - boolean whether backend has no matching results

An additional file db.json is provided which contains the set of values for each slot.

Note: The date values in the dataset are normalized as the constants, "base_date_plus_X", for X from 0 to 6. X=0 corresponds to the current date (i.e. 'today'), X=1 is 'tomorrow', etc. This is done to allow handling of relative references to dates (e.g. 'this weekend', 'next Wednesday', etc). The parsing of such phrases should be done as a separate pre-processing step.