Implicit Hate Speech Detection (ToxiGen)

# Implicit Hate Speech Detection (ToxiGen)
# Based on Hartvigsen et al., ACL 2022
# Paper: https://aclanthology.org/2022.acl-long.234/
# Dataset: https://github.com/microsoft/TOXIGEN
#
# This task focuses on detecting implicit hate speech -- statements that
# are toxic toward demographic groups but do not use overt slurs or
# explicitly hateful language. The ToxiGen dataset uses large language
# models to generate adversarial examples that challenge toxicity classifiers.
#
# Toxicity Labels:
# - TOXIC: The statement expresses negativity, stereotypes, or prejudice toward a group
# - BENIGN: The statement is neutral or positive; does not demean any group
#
# Implicitness Labels:
# - EXPLICIT: Uses overt slurs, direct insults, or clearly hateful language
# - IMPLICIT: Conveys negativity through stereotypes, dog whistles, or coded language
# - BORDERLINE: Ambiguous; could be interpreted either way depending on context
# - NEUTRAL: Not applicable (for benign statements)
#
# Annotation Guidelines:
# 1. Read the statement carefully without assuming intent
# 2. Consider whether the statement promotes negative stereotypes or prejudice
# 3. Implicit hate may use seemingly neutral language to convey harmful ideas
# 4. Consider the target group and historical context of statements
# 5. A statement can be factually true but still be used in a toxic framing
# 6. Look for generalizations about groups, dehumanizing comparisons, or dog whistles
# 7. Benign statements may mention demographic groups without negativity
# 8. Provide clear reasoning for your toxicity and implicitness decisions
#
# Important Considerations:
# - Some statements are designed to be adversarial (hard to classify)
# - Context matters: the same words can be toxic or benign depending on framing
# - Avoid over-sensitivity: mentioning a group is not inherently toxic
# - Avoid under-sensitivity: lack of slurs does not mean lack of toxicity

annotation_task_name: "Implicit Hate Speech Detection (ToxiGen)"
task_dir: "."

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

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

annotation_schemes:
  # Step 1: Toxicity classification
  - annotation_type: radio
    name: toxicity_label
    description: "Is this statement toxic (harmful, stereotyping, prejudiced) or benign (neutral, positive)?"
    labels:
      - "toxic"
      - "benign"
    keyboard_shortcuts:
      "toxic": "t"
      "benign": "b"
    tooltips:
      "toxic": "The statement expresses negativity, promotes stereotypes, or conveys prejudice toward a demographic group"
      "benign": "The statement is neutral or positive and does not demean or stereotype any group"

  # Step 2: Implicitness level
  - annotation_type: radio
    name: implicitness
    description: "If toxic, how implicit is the hate speech? If benign, select neutral."
    labels:
      - "explicit"
      - "implicit"
      - "borderline"
      - "neutral"
    tooltips:
      "explicit": "Uses overt slurs, direct insults, or clearly hateful language targeting a group"
      "implicit": "Conveys toxicity through stereotypes, dog whistles, coded language, or subtle framing without overt slurs"
      "borderline": "Ambiguous -- could be toxic or benign depending on interpretation and context"
      "neutral": "Not applicable (statement is benign)"

  # Step 3: Target group identification
  - annotation_type: radio
    name: target_group
    description: "Which demographic group is targeted by the statement (if any)?"
    labels:
      - "racial-minority"
      - "religious-group"
      - "gender-identity"
      - "sexual-orientation"
      - "disability"
      - "nationality"
      - "political-group"
      - "other"
      - "none"
    tooltips:
      "racial-minority": "Statement targets a racial or ethnic minority group"
      "religious-group": "Statement targets a religious group"
      "gender-identity": "Statement targets based on gender identity or expression"
      "sexual-orientation": "Statement targets based on sexual orientation"
      "disability": "Statement targets people with disabilities"
      "nationality": "Statement targets based on nationality or immigration status"
      "political-group": "Statement targets a political group or affiliation"
      "other": "Statement targets a group not listed above"
      "none": "No group is targeted (benign statement)"

  # Step 4: Explanation
  - annotation_type: text
    name: explanation
    description: "Explain why the text is toxic or benign. If toxic, describe the implicit cues or stereotypes present."

annotation_instructions: |
  You will be shown statements that may or may not contain implicit hate speech targeting various demographic groups. Your task is to:
  1. Classify the statement as TOXIC or BENIGN.
  2. If toxic, assess how implicit the hate is (explicit slurs vs. coded language vs. borderline).
  3. Identify the demographic group being targeted, if any.
  4. Explain your reasoning, pointing out specific cues in the text.

  Key points:
  - Implicit hate speech avoids slurs but promotes stereotypes or prejudice through subtle framing.
  - A statement mentioning a group is NOT automatically toxic.
  - Consider how the statement might affect members of the referenced group.
  - Some statements are intentionally adversarial and difficult to classify.

html_layout: |
  <div style="padding: 15px; max-width: 800px; margin: auto;">
    <div style="background: #fef3c7; border: 1px solid #f59e0b; border-radius: 8px; padding: 8px 12px; margin-bottom: 16px; font-size: 13px; color: #92400e;">
      <strong>Content Warning:</strong> This task may contain offensive or hateful language for research purposes.
    </div>
    <div style="font-size: 16px; line-height: 1.7; padding: 16px; background: #f8fafc; border: 1px solid #e2e8f0; border-radius: 8px;">
      {{text}}
    </div>
    <div style="margin-top: 10px; font-size: 13px; color: #6b7280;">
      <em>Generation method: {{generation_method}} | Target category: {{target_demographic_category}}</em>
    </div>
  </div>

allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 3
allow_skip: true
skip_reason_required: false