Structured Fact Verification (FEVEROUS)

# Structured Fact Verification (FEVEROUS)
# Based on Aly et al., NeurIPS 2021
# Paper: https://arxiv.org/abs/2106.05707
# Dataset: https://fever.ai/dataset/feverous.html
#
# This task requires annotators to verify claims against Wikipedia pages
# containing a mix of unstructured text paragraphs, tables, lists, and
# infoboxes. Unlike simpler fact verification, FEVEROUS demands reasoning
# across multiple evidence modalities simultaneously.
#
# Verdict Labels:
# - SUPPORTS: The page content provides evidence that the claim is true
# - REFUTES: The page content provides evidence that contradicts the claim
# - NOT ENOUGH INFO: The page does not contain sufficient evidence
#
# Evidence Types:
# - text-evidence: Sentences from paragraphs that support/refute the claim
# - table-evidence: Cells or rows from tables containing relevant data
# - list-evidence: Items from bulleted or numbered lists
# - infobox-evidence: Fields from the page's infobox/sidebar
#
# Annotation Guidelines:
# 1. Read the claim carefully and identify all sub-claims within it
# 2. Scan the entire page including text, tables, lists, and infobox
# 3. For each sub-claim, locate the most relevant evidence
# 4. Evidence may come from multiple modalities (e.g., text + table)
# 5. Mark the evidence type for each highlighted span
# 6. Assess whether the combined evidence is sufficient
# 7. Claims requiring information NOT on the page are NOT ENOUGH INFO
# 8. Provide a justification that references specific evidence pieces
#
# Important Notes:
# - A claim can be REFUTED even if some parts are true
# - Multiple evidence types may be needed for a single claim
# - NOT ENOUGH INFO means the page lacks relevant information entirely

annotation_task_name: "Structured Fact Verification (FEVEROUS)"
task_dir: "."

data_files:
  - sample-data.json
item_properties:
  id_key: "id"
  text_key: "claim"

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

annotation_schemes:
  # Step 1: Overall verdict
  - annotation_type: radio
    name: verdict
    description: "Based on the page content (text, tables, lists), is the claim supported, refuted, or is there not enough information?"
    labels:
      - "supports"
      - "refutes"
      - "not-enough-info"
    keyboard_shortcuts:
      "supports": "s"
      "refutes": "r"
      "not-enough-info": "n"
    tooltips:
      "supports": "The page provides sufficient evidence that the claim is true"
      "refutes": "The page provides evidence that contradicts the claim"
      "not-enough-info": "The page does not contain enough information to verify or refute"

  # Step 2: Evidence type highlighting
  - annotation_type: span
    name: evidence_type
    description: "Highlight evidence in the page content and classify by source type"
    labels:
      - "text-evidence"
      - "table-evidence"
      - "list-evidence"
      - "infobox-evidence"
    label_colors:
      "text-evidence": "#3b82f6"
      "table-evidence": "#22c55e"
      "list-evidence": "#f59e0b"
      "infobox-evidence": "#a855f7"
    tooltips:
      "text-evidence": "Evidence from paragraph text on the page"
      "table-evidence": "Evidence from table cells or rows"
      "list-evidence": "Evidence from bulleted or numbered lists"
      "infobox-evidence": "Evidence from the page infobox or sidebar"
    allow_overlapping: false

  # Step 3: Evidence sufficiency
  - annotation_type: radio
    name: evidence_sufficiency
    description: "How sufficient is the evidence on this page for making a verdict?"
    labels:
      - "sufficient"
      - "partial"
      - "insufficient"
    tooltips:
      "sufficient": "The page contains all evidence needed to conclusively determine the verdict"
      "partial": "The page contains some relevant evidence but not enough for full certainty"
      "insufficient": "The page lacks meaningful evidence related to the claim"

  # Step 4: Justification
  - annotation_type: text
    name: justification
    description: "Explain your verdict and describe which evidence (text, table, list, infobox) you relied on."

annotation_instructions: |
  You will be shown a claim and a Wikipedia page that contains text paragraphs, tables, lists, and/or infoboxes. Your task is to:
  1. Determine whether the page content SUPPORTS, REFUTES, or does NOT provide ENOUGH INFO for the claim.
  2. Highlight evidence spans and classify them by type (text, table, list, or infobox).
  3. Assess whether the evidence you found is sufficient, partial, or insufficient.
  4. Write a justification explaining your verdict and referencing the specific evidence you found.

  Key considerations:
  - Evidence may span multiple modalities (e.g., a name in text and a date in a table).
  - A claim with partially correct information may still be REFUTED if key details are wrong.
  - If the page simply does not address the claim, choose NOT ENOUGH INFO.

html_layout: |
  <div style="padding: 15px; max-width: 900px; margin: auto;">
    <div style="background: #fffbeb; border-left: 4px solid #f59e0b; padding: 12px 16px; border-radius: 4px; margin-bottom: 20px;">
      <strong style="color: #92400e;">Claim:</strong>
      <span style="font-size: 16px; line-height: 1.6;">{{claim}}</span>
    </div>
    <h3 style="color: #1e40af; margin-bottom: 10px; border-bottom: 2px solid #dbeafe; padding-bottom: 6px;">{{page_title}}</h3>
    <div style="font-size: 15px; line-height: 1.7; border: 1px solid #e5e7eb; border-radius: 8px; padding: 16px; background: #f9fafb;">
      {{page_content}}
    </div>
  </div>

allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 2
allow_skip: true
skip_reason_required: false