KGConv, a Conversational Corpus grounded in Wikidata

Dataset Summary

KGConv is a large corpus of 71k english conversations where each question-answer pair is grounded in a Wikidata fact. The conversations were generated automatically: in particular, questions were created using a collection of 10,355 templates; subsequently, the naturalness of conversations was improved by inserting ellipses and coreference into questions, via both handcrafted rules and a generative rewriting model. The dataset thus provides several variants of each question (12 on average), organized into 3 levels of conversationality. KGConv can further be used for other generation and analysis tasks such as single-turn question generation from Wikidata triples, question rewriting, question answering from conversation or from knowledge graphs and quiz generation.

Languages

English.

Dataset Structure

The dataset has three components:

• conversation configs, divided in several themes that correspond to configs of the form conversations.theme, where theme has to be replaced by one of the following: country, food, historical_event, ideology, molecular_entity, person, space_object, taxon, with_unseen_properties;
• labels, a config that contains labels for all entities and properties involved in the conversations;
• templates, a config that contains the templates that where used for generating questions.

Data Instances

Instance from the configs with name of the form "conversations.theme" (e.g. "conversations.country") have the following form:

{
    "conversation_id": "69795",
    "root_neighbourhood": [
      [
        "Q6138903",
        "P106",
        "Q82955"
      ],
      [
        "Q6138903",
        "P19",
        "Q3408680"
      ],
      ...
    ],
    "conversation": [
      {
        "triple": [
          "Q691",
          "P30",
          "Q538"
        ],
        "question variants": [
          {
            "out-of-context": "In which continent is Papua New Guinea located?",
            "in-context": "In which continent is Papua New Guinea located?",
            "in-context subject ref": "Papua New Guinea",
            "synthetic-in-context": "In which continent is Papua New Guinea located?"
          },
          {
            "out-of-context": "In what continent is Papua New Guinea in?",
            "in-context": "In what continent is Papua New Guinea in?",
            "in-context subject ref": "Papua New Guinea",
            "synthetic-in-context": "In what continent is Papua New Guinea in?"
          },
          ...
        ],
        "answer": "Oceania"
      },
      {
        "triple": [
          "Q691",
          "P38",
          "Q200759"
        ],
        "question variants": [
          {
            "out-of-context": "What is accepted as the currency of Papua New Guinea?",
            "in-context": "What is accepted as the currency of Papua New Guinea?",
            "in-context subject ref": "Papua New Guinea",
            "synthetic-in-context": "What is accepted as the currency?"
          },
          {
            "out-of-context": "What is the currency of Papua New Guinea?",
            "in-context": "What is the currency of Papua New Guinea?",
            "in-context subject ref": "Papua New Guinea",
            "synthetic-in-context": "What is the currency?"
          },
          ...
        ],
        "answer": "kina"
      },
      ...

Instances from the labels config are like this:

{
    "entity": "Q39",
    "labels": [
      "Swiss Confederation",
      "CHE",
      "Confoederatio Helvetica",
      "Swiss",
      "Schweiz",
      "SUI",
      "Switzerland",
      "CH",
      "Suisse",
      "Svizzera"
    ],
    "preferred_label": "Switzerland"
}

Instances from the templates config are as follows.

{
  "template_key": {
    "p": "P1201",
    "s_types": [
      "Q149918"
    ],
    "o_types": []
  },
  "templates": [
    {
      "left": "what is the space tug of ",
      "right": "?",
      "source": "interface:automatic labeler"
    },
    {
      "left": "what was the space tug of ",
      "right": "?",
      "source": "interface:624dc1cd4432b5035ba082df"
    },
    ...
  ]
}

Data Fields

The fields from the configs with name of the form "conversations.theme" (e.g. "conversations.country") are the following:

• conversation: list of dicts; each dict reprensent one question+answer and has the following fields:
  • conversation_id: string
  • root_neighbourhood: list of triples (each triple is itself represented by a list of 3 string elements) that constitute the neighbourhood of the conversation root entity in the knowledge graph (see the LREC publication for more details)
  • triple: triple on which the question is based (list of three string elements)
  • question variants: list of dict; each dict contain several forms of a question obtained via a given template (see the LREC publication for more details)
    • out-of-context: one form of the question variant
    • in-context: another form of the question variant
    • in-context subject ref: how the subject is referred to in the in-context form
    • synthetic-in-context: yet another form of the question variant
  • answer: answer to the question (string)

The fields from the labels config are the following:

• entity: string, id of the entity
• labels: list of strings
• preferred_label: string

The fields from the templates config are the following:

• template_key: a dict containing the conditions for using the templates listed in templates, with the following fields:
  • p: id of the property
  • s_types: required types for subject
  • o_types: require types for object
• templates: list of dicts representing templates; each dict has the following fields:
  • left: left part of the template
  • right: right part of the template
  • source: origin of the template (string)

Additional Information

Licensing Information

This software is distributed under the Creative Commons Attribution 4.0 International, the text of which is available at https://spdx.org/licenses/CC-BY-4.0.html or see the "license.txt" file for more details.

Citation Information

@article{brabant2023kgconv,
      title={KGConv, a Conversational Corpus grounded in Wikidata}, 
      author={Quentin Brabant and Gwenole Lecorve and Lina M. Rojas-Barahona and Claire Gardent},
      year={2023},
      eprint={2308.15298},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}