Skip to content

Latest commit

 

History

History
73 lines (50 loc) · 3.2 KB

README.md

File metadata and controls

73 lines (50 loc) · 3.2 KB
Raw

DSTC9 Track 1 Dataset

This directory contains the official training/validation datasets for DSTC9 Track 1. The evaluation dataset details can be found from data_eval/.

If you want to publish experimental results with this dataset or use the baseline models, please cite the following article:

@article{kim2020domain,
  title={Beyond Domain APIs: Task-oriented Conversational Modeling with Unstructured Knowledge Access},
  author={Seokhwan Kim and Mihail Eric and Karthik Gopalakrishnan and Behnam Hedayatnia and Yang Liu and Dilek Hakkani-Tur},
  journal={arXiv preprint arXiv:2006.03533}
  year={2020}
}

NOTE: This paper describes an earlier version of the data, which differs from the official challenge dataset in detail.

Data

We are releasing the data divided into the following two subsets:

The ground-truth labels for Knowledge Selection task refer to the knowledge snippets in knowledge.json.

Participants will develop systems to take logs.json as an input and generates the outputs following the same format as labels.json.

JSON Data Formats

Log Objects

The logs.json file includes the list of instances each of which is a partial conversation from the beginning to the target user turn. Each instance is a list of the following turn objects:

  • speaker: the speaker of the turn (string: "U" for user turn/"S" for system turn)
  • text: utterance text (string)

Label Objects

The labels.json file provides the ground-truth labels and human responses for the final turn of each instance in logs.json. It includes the list of the following objects in the same order as the input instances:

  • target: whether the turn is knowledge-seeking or not (boolean: true/false)
  • knowledge: [
    • domain: the domain identifier referring to a relevant knowledge snippet in knowledge.json (string)
    • entity_id: the entity identifier referring to a relevant knowledge snippet in knowledge.json (integer for entity-specific knowledge or string "*" for domain-wide knowledge)
    • doc_id: the document identifier referring to a relevant knowledge snippet in knowledge.json (integer) ]
  • response: knowledge-grounded system response (string)

NOTE: knowledge and response exist only for the target instances with true for the target value.

Knowledge Objects

The knowledge.json contains the unstructured knowledge sources to be selected and grounded in the tasks. It includes the domain-wide or entity-specific knowledge snippets in the following format:

  • domain_id: domain identifier (string: "hotel", "restaurant", "train", "taxi", etc.)
    • entity_id: entity identifier (integer or string: "*" for domain-wide knowledge)
      • name: entity name (string; only exists for entity-specific knowledge)
      • docs
        • doc_id: document identifier (integer)
          • title: document title (string)
          • body: document body (string)