dataset

The Missci Dataset

We provide a validation split and a test split for the Missci dataset to allow prompt engineering without compromising the unseen test set. Both files are provided as .jsonl format.

Argument Structure in the Missci Dataset

Each line in the files constitutes one argument. Each argument is provided using JSON formatting as below:

{
    "id": "<a unique identifier>",
    "meta": {
        "fc_article_url": "<some information about the fact-checking article such as date, url, ...>",
        "...": "..."
    },
    "argument": {
        "claim": "<The claim of the argument that misrepresents a scientific publication>",
        "hidden_premises": [
            "<optional hidden premises that are accurate but not part of the scientific publication>"
        ],
        "accurate_premise_p0": {
            "premise": "<The accurate premise p0>"
        },
        "fallacies": [
            {
                "consolidation_comment": "<annotator's summary of the fallacies>",
                "fallacy_context": "<fallacy context si. This may be an empty string if p0 contains all the information to detect the fallacies>",
                "id": "<unique identifier for this fallacy, including all interchangeable fallacies>",
                "interchangeable_fallacies": [
                    {
                        "premise": "<fallacious premise 1>",
                        "class": "<applied fallacy class 1>",
                        "id": "<unique identifier for this interchangeable fallacy>"
                    },
                    {
                        "...": "<optionally additional interchangeable fallacies>"
                    }
                ]
            },
            {
                "...": "<optionally additional fallacies>"
            }
        ]
    },
    "study": {
        "url": "<url to the misrepresented scientific publication>"
    }
}

Each argument contains

• a single inaccurate claim based on
• a single accurate premise $p_0$ and
• at least one fallacy $F_i$ containing a fallacy context $s_i$ and at least one interchangeable fallacy of
  • a fallacious premise
  • the applied fallacy class.