Skip to content

Latest commit

 

History

History
163 lines (138 loc) · 6.5 KB

data.md

File metadata and controls

163 lines (138 loc) · 6.5 KB
Raw

SciFact-Open data

After running get_data.sh as described in the README, a data folder will populate, containing the following files:

  • claims.jsonl: Claims in SciFact-Open, annotated with evidence.
  • claims_metadata.jsonl: Metadata associated with each claim.
  • corpus.jsonl: The full SciFact-Open corpus of 500K research abstracts from S2ORC.
  • corpus_candidates.jsonl: Subset of documents from corpus.jsonl that were retrieved for at least one claim.

Detailed descriptions of each file are included below.

claims.jsonl

A .jsonl file with one claim per line, annotated with evidence. The schema is as follows:

{
    "id": number,                   # An integer claim ID.
    "claim": string,                # The text of the claim.
    "evidence": {                   # The evidence for the claim.
        [doc_id]:                  # The rationales for a single document, keyed by S2ORC ID.
            {
                "provenance": enum("citation" | "pooling")    # Evidence source. See below.
                "label": enum("SUPPORT" | "CONTRADICT"),      # Veracity label.
                "sentences": number[],                        # Evidence "highlights".
                "model_ranks": {[model_name]: number} | None  # Ranks for pooled evidence.
            }
    }
}

Note that this format is altered slightly from the format for SciFact-Orig; see below for details.

A few of these fields warrant explanation.

  • provenance: As described in the paper, evidence in SciFact-Open comes from two sources (see Sec. 3 of the paper for more details):
    1. Evidence from SciFact-Orig, obtained via citation links.
    2. Evidence added in SciFact-Open via pooling. The provenance field indicates the source for each piece of evidence in the final SciFact-Open dataset.
  • sentences: Evidence "highlights" justifying each labeling decision. The source of these highlights is different for citation evidence vs. pooling evidence:
    • For citation evidence, highlights were hand-annotated by experts and are high-quality.
    • For pooling evidence, highlights were not hand-annotated. Instead, we include the predicted highglights from the model(s) that identified the evidence. These predictions may be incorrect, but are included for convenience and to facilitate future research.
  • model_ranks: For pooled data collection, predicted evidence is ranked by model confidence, and the k most-confidence predictions for each model are annotated; we used k=250.
    • For pooling evidence, this field provides the rank assigned to this piece of evidence by each model.
    • For citation evidence, this field is set to None.

Below, we provide an example of a claim with two pieces of evidence: one from SciFact-Orig, and one identified via pooling.

{
  "id": 170,
  "claim": "Bariatric surgery reduces resolution of diabetes.",
  "evidence": {
    "5824985": {
      "provenance": "citation",
      "label": "CONTRADICT",
      "sentences": [
        10,
        15
      ],
      "model_ranks": None
    },

    "7854739": {
      "provenance": "pooling",
      "label": "CONTRADICT",
      "sentences": [
        2,
        5,
        6
      ],
      "model_ranks": {
        "multivers_10": 4556,
        "multivers_20": 3881,
        "paragraph_joint": 947,
        "vert5erini": 120
      }
    }
  }
}

Differences from SciFact-Orig

The format here is slightly different from the format for the original SciFact dataset. In that work, a single document could be associated with multiple evidence sets, each of which provides sufficient justification to support or refute the claim. For instance:

{
  # SciFact-Orig
  "id": 263,
  "claim": "Citrullinated proteins externalized in neutrophil extracellular traps act indirectly to disrupt the inflammatory cycle.",
  "evidence": {
    "30041340": [   # Multiple evidence sets associated with a single document.
      { "sentences": [0, 1],
        "label": "CONTRADICT" },
      { "sentences": [11],
        "label": "CONTRADICT" }
    ]
  }
}

However, we didn't end up using this evidence set structure in our modeling or evaluation. Therefore, in this work, we simplify things a bit by "flattening" evidence sets, as follows:

{
  # SciFact-Open
  "id": 263,
  "claim": "Citrullinated proteins externalized in neutrophil extracellular traps act indirectly to disrupt the inflammatory cycle.",
  "evidence": {
    "30041340": {   # Flattened evidence, including all highlights for this document.
        "sentences": [0, 1, 11],
        "label": "CONTRADICT"
    },
  }
}

claims_metadata.jsonl

Each line includes metadata for a single claim, ordered as in claims.jsonl. This information isn't needed for any modeling task related to SciFact-Open, but may be of interest for those interested in designing and analyzing datasets. As background, claims in SciFact were created by re-writing citation sentences found in documents in the S2ORC corpus. We refer to the document used as the source of the claim as the "source document".

{
  "id": number   # Same ID's as in `claims.jsonl`.
  "source_doc_id": number # The S2ORC ID of the document
  "source_metadata": {    # Metadata on claim source document.
    "paper_id": string
    "title": string
    "year": number
    "arxiv_id": string | None
    "pmc_id": string | None
    "pubmed_id": string | None
    "venue": string | None
    "journal": string | None
  },
  "cited_doc_ids": [number] # Documents mentioned in source citation
}

corpus.jsonl

A single line represents a document from the SciFact-Open corpus.

{
    "doc_id": number,               # The document's S2ORC ID.
    "title": string,                # The title.
    "abstract": string[],           # The abstract, written as a list of sentences.
    "metadata": {
        "paper_id": string,         # Redundant with `doc_id` above.
        "title": string,
        "year": number,
        "venue": string,
        "s2orc_version: string
    },
    "scifact_orig": bool            # True if this document was part of the corpus from SciFact-Orig
}

corpus_candidates.jsonl

This file contains the subset of ~12K documents from corpus.jsonl that either contain evidence indicated via citation links (from SciFact-Orig), or were identified via pooling as potential evidence candidates (from SciFact-Open).