classicvalues / ml-tree-dst

Conversational Semantic Parsing for Dialog State Tracking

We introduce TreeDST (Tree-based Dialog State Tracking), a multi-turn, multi-domain task-oriented dialog dataset annotated with tree-based user dialog states and system dialog acts. The goal of this dataset is to provide a novel solution for end-to-end dialog state tracking as a conversational semantic parsing task. Please refer to our paper for Conversational Semantic Parsing for Dialog State Tracking for details.

Task Description

The task in TreeDST is to predict the user dialog state for each turn of a conversation. The dialog state is a representation of the user's goal up to the current turn of the conversation.

Dataset Description

The dataset contains 27,280 conversations covering 10 domains with shared types of person, time and location. The dataset and schema can be accessed in the dataset folder. A tool for visualizing the data is in the "dotted" format is provided at utils folder.

License

The code in this repository is licensed according to the LICENSE file.

The TreeDST dataset is licensed under the Creative Commons Attribution-ShareAlike 3.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by-sa/3.0/.

Declaration

This dataset contains computer-generated synthetic conversation flows that has been human annotated to provide richer and more natural dialogue. This dataset does not include any data from Siri production users.