Social Bias Frames (SBIC)

# Social Bias Frames (SBIC)
# Based on Sap et al., ACL 2020
# Paper: https://aclanthology.org/2020.acl-main.409/
# Dataset: https://maartensap.com/social-bias-frames/
#
# Social Bias Frames capture multiple dimensions of biased language:
# 1. Offensiveness: Is the post offensive?
# 2. Intent: Was the speaker trying to be offensive?
# 3. Lewdness: Is it sexual or lewd?
# 4. Group implication: Does it imply something about a group?
# 5. Targeted group: Which group is targeted?
# 6. Implied statement: What stereotype is implied?
#
# Annotation Guidelines:
# 1. Consider the post in isolation (not broader context)
# 2. Offensiveness is about potential to offend, not your personal reaction
# 3. Intent considers whether offense seems deliberate
# 4. Group implications apply when stereotypes about demographics are implied
# 5. Targeted groups should be specific (e.g., "Black women" not just "women")
# 6. Implied statements should be written as generalizations
#
# Key Distinctions:
# - A post can be offensive without intent (ignorance)
# - A post can have intent without being offensive (failed joke)
# - Group implications require stereotyping, not just mentioning groups

annotation_task_name: "Social Bias Frames"
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: Offensiveness
  - annotation_type: radio
    name: offensive
    description: "Could this post be considered offensive to anyone?"
    labels:
      - "Yes, offensive"
      - "Maybe offensive"
      - "Not offensive"
    tooltips:
      "Yes, offensive": "The post would likely offend members of some group"
      "Maybe offensive": "The post could be seen as offensive by some people"
      "Not offensive": "The post is unlikely to offend anyone"

  # Step 2: Intent
  - annotation_type: radio
    name: intent
    description: "Was the intent of the post to be offensive or hurtful?"
    labels:
      - "Intentionally offensive"
      - "Possibly intentional"
      - "Not intentional"
      - "Can't tell"
    tooltips:
      "Intentionally offensive": "The speaker clearly meant to offend or demean"
      "Possibly intentional": "The offense might have been deliberate"
      "Not intentional": "The speaker likely didn't mean to offend (ignorance, poor wording)"
      "Can't tell": "Intent is unclear from the post"

  # Step 3: Lewdness
  - annotation_type: radio
    name: lewd
    description: "Is the post lewd or sexual in nature?"
    labels:
      - "Yes"
      - "Somewhat"
      - "No"
    tooltips:
      "Yes": "The post is explicitly sexual or lewd"
      "Somewhat": "The post has sexual undertones or innuendo"
      "No": "The post is not sexual"

  # Step 4: Group implication
  - annotation_type: radio
    name: group_implication
    description: "Does the post imply something negative about a group of people?"
    labels:
      - "Yes, implies stereotype"
      - "Mentions group but no stereotype"
      - "No group mentioned"
    tooltips:
      "Yes, implies stereotype": "The post implies a generalization or stereotype about a demographic group"
      "Mentions group but no stereotype": "A group is mentioned but no stereotype is implied"
      "No group mentioned": "No demographic group is referenced"

  # Step 5: Target group (if applicable)
  - annotation_type: multiselect
    name: target_group
    description: "Which group(s) are targeted? (Select all that apply)"
    labels:
      - "Women"
      - "Men"
      - "Black people"
      - "Asian people"
      - "Hispanic/Latino people"
      - "White people"
      - "LGBTQ+ people"
      - "Muslims"
      - "Jewish people"
      - "Immigrants"
      - "Disabled people"
      - "Elderly people"
      - "Poor/working class"
      - "Other group"
    min_selections: 0
    max_selections: 14

allow_all_users: true
instances_per_annotator: 100
annotation_per_instance: 3
allow_skip: true
skip_reason_required: false