Dataset Card for Canard_Wiki-augmented

Summary

This is a dataset of fact-retrieving conversations about Wikipedia articles, with all responses grounded in a specific segment of text in the referenced Wikipedia article. It is an extended version of Canard and QuAC datasets, augmented with the contexts of English Wikipedia.

Supported Tasks

The dataset is intended to train a factually-consistent conversational model able to ground all its responses to the corresponding source(s). However, the data can also be used to evaluate the information retrieval (IR) system for given queries, for contextual disambiguation of the queries from a conversation, etc.

Dataset Structure

The dataset can be loaded by simply choosing a split (train or test) and calling:

import datasets
canard_augm_test = datasets.load_dataset("gaussalgo/Canard_Wiki-augmented", split="test")

print(canard_augm_test[0])  # print the first sample

Data Instances

The samples of Canard_Wiki-augmented have this format:

{'History': ['Anna Politkovskaya', 'The murder remains unsolved, 2016'],
 'QuAC_dialog_id': 'C_0aaa843df0bd467b96e5a496fc0b033d_1',
 'Question': 'Did they have any clues?',
 'Question_no': 1,
 'answer': 'Her colleagues at Novaya gazeta protested that until the instigator or sponsor of the crime was identified, arrested and prosecuted the case was not closed.'
 'Rewrite': 'Did investigators have any clues in the unresolved murder of Anna Politkovskaya?',
 'true_page_title': 'Anna Politkovskaya',
 'true_contexts': 'In September 2016 Vladimir Markin, official spokesman for (...)',
 'true_contexts_wiki': 'Anna Stepanovna Politkovskaya was a US-born Russian journalist (...)',
 'extractive': True
 'retrieved_contexts': ['Clues was an indie rock band from Montreal, Canada formed by Alden Penner (...)',
                        'High Stakes is a British game show series hosted by Jeremy Kyle, in which (...)']

Data Fields

• History: History of the conversation from Canard. The first two entries of the conversation are always synthetic.
• QuAC_dialog_id: Dialogue ID mapping the conversation to the original QuAC dataset (dialogue_id in QuAC).
• Question: Current question of the user from Canard.
• Question_no: Ordering of the user's question from the conversation, originally from Canard.
• answer: Correctly extracted answer to a given question from a relevant Wikipedia article (true_contexts). Note that some of the questions are open, thus the listed answer is not the only correct possibility.
• Rewrite: A rephrased version of Question, manually disambiguated from the context of History by the annotators of Canard.
• true_page_title: Title of the Wikipedia article containing answer. wikipedia_page_title from QuAC.
• true_contexts: An excerpt of the paragraph with an answer from the Wikipedia article titled true_page_title.
• true_contexts_wiki: A full contents of Wikipedia article (text from Wikipedia dataset), where true_page_title matches Wikipedia title. Note that the Wikipedia dataset was retrieved on 2nd of April, 2023.
• extractive: A flag whether the answer in this sample can be found as an exact-match in true_contexts_wiki.
• retrieved_contexts: "Distractor" contexts retrieved from the full Wikipedia dataset using the okapi-BM25 IR system on a Rewrite question.

Data Splits

• train split is aligned with the training splits of Canard and QuAC.
• test split matches the validation split of QuAC and the test split of Canard (where the conversation ids match).

Licensing

This dataset is composed of QuAC (MIT), Canard (CC BY-SA 4.0) and Wikipedia (CC BY SA 3.0). Canard_Wiki-augmented is therefore licensed under CC BY-SA 4.0 as well, allowing it to be also commercially used.

Cite

If you use this dataset in a research, do not forget to cite the authors of original datasets, that Canard_Wiki-augmented is derived from: QuAC, Canard.