FactSpan: Multilingual Fact-Checking Dataset

The FactSpan dataset is an extension of the X-Fact dataset, designed to support multilingual fact-checking research. This dataset overcomes limitations in existing datasets by incorporating recent data from the ClaimReview Markup for Data Commons Feed and providing detailed annotations.

Key Features:

Data Source: Claims are sourced from both the X-Fact dataset (up to 2020) and the Data Commons Feed (post-2020).

Validity: Claims are filtered to include only those from organizations recognized by the International Fact-Checking Network (IFCN) and Duke Reporters’ Lab, ensuring high reliability.

Standardized Labels: Verdict labels are standardized into five categories: False, Mostly False, Partly False/Misleading, Mostly True, and True.

Annotations (Annotated Dataset Only): The FactSpan_annotated.csv dataset includes rich annotations generated using GPT-3.5:

label: The standardized verdict label.

claim: The fact-checked claim.

claimDate: The date of the claim.

claim_year: The year of the claim.

language: The language of the claim.

Position Statements: Indicates the presence of position statements.

Entity/Event Properties: Indicates the presence of entity or event properties.

Quote: Indicates the presence of quotes.

Numerical Data: Indicates the presence of numerical data.

claim type: Categorizes the claim as factual or opinion.

topics: Categorizes the claim into one of five predefined topics (Health and Pandemics, Politics and Governance, Society and Culture, Economy and Environment, Conflict and Security).

mapped_label: An additional mapped label, for edge cases or further label mappings.

Unannotated Dataset: The FactSpan.csv dataset includes:

label: The standardized verdict label.

claim: The fact-checked claim.

claimDate: The date of the claim.

language: The language of the claim.

Purpose:

This dataset aims to facilitate research in multilingual fact-checking, providing a comprehensive and up-to-date resource for developing and evaluating fact-checking models.

Repository:

The dataset is maintained in the GitHub repository. The repository also contains scripts for expanding and updating the dataset.