Implicit Hate Speech Detection

# Implicit Hate Speech Detection
# Based on ElSherief et al., EMNLP 2021 "Latent Hatred"
# Paper: https://aclanthology.org/2021.emnlp-main.29/
# Dataset: https://github.com/SALT-NLP/implicit-hate
#
# Implicit hate speech is harder to detect than explicit hate because it
# relies on stereotypes, coded language, and indirect expression.
#
# Six-Category Taxonomy:
# 1. Grievance (24.2%): Positions majority groups as unfairly disadvantaged
# 2. Incitement (20.0%): Promotes hate groups/ideologies, flaunts in-group power
# 3. Stereotypes (17.9%): Associates groups with negative attributes via euphemisms
# 4. Inferiority (13.6%): Implies some group is of lesser value
# 5. Irony (12.6%): Uses sarcasm, humor, satire to demean
# 6. Threats (10.5%): Indirect threats to safety, well-being, reputation
#
# Annotation Guidelines:
# 1. First determine if the text contains implicit hate (vs explicit or none)
# 2. Implicit hate lacks slurs but conveys hateful meaning through context
# 3. Consider: Would someone from the target group feel attacked?
# 4. Look for coded language, dog whistles, and stereotypical references
# 5. Irony/sarcasm requires understanding the speaker's actual intent
#
# Key Distinction from Explicit Hate:
# - Explicit: Uses slurs, direct insults, clear threats
# - Implicit: Uses stereotypes, innuendo, coded language, sarcasm

annotation_task_name: "Implicit Hate Speech Detection"
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: Hate classification
  - annotation_type: radio
    name: hate_type
    description: "Does this text contain hate speech?"
    labels:
      - "Explicit hate"
      - "Implicit hate"
      - "Not hateful"
    tooltips:
      "Explicit hate": "Contains clear slurs, direct insults, or obvious threats"
      "Implicit hate": "Conveys hate through stereotypes, coded language, or indirect means"
      "Not hateful": "Does not contain hateful content"

  # Step 2: Implicit hate category (if implicit)
  - annotation_type: radio
    name: implicit_category
    description: "What type of implicit hate is expressed?"
    labels:
      - "Grievance"
      - "Incitement"
      - "Stereotypes"
      - "Inferiority"
      - "Irony"
      - "Threats"
    tooltips:
      "Grievance": "Positions majority groups as unfairly disadvantaged (e.g., 'reverse racism', 'they get special treatment')"
      "Incitement": "Promotes hate groups/ideologies, celebrates in-group power (e.g., praising extremist symbols)"
      "Stereotypes": "Associates groups with negative attributes using euphemisms or coded language"
      "Inferiority": "Implies some group or person is of lesser value or capability"
      "Irony": "Uses sarcasm, humor, or satire to mock or demean a group"
      "Threats": "Indirect threats to body, well-being, reputation, or rights"

  # Step 3: Target group
  - annotation_type: radio
    name: target
    description: "Which group is targeted?"
    labels:
      - "Racial/ethnic minority"
      - "Religious group"
      - "Women"
      - "LGBTQ+ people"
      - "Immigrants"
      - "Other group"
      - "No specific group"
    tooltips:
      "Racial/ethnic minority": "Black, Asian, Hispanic, Indigenous, or other racial/ethnic groups"
      "Religious group": "Muslims, Jewish people, or other religious groups"
      "Women": "Women or girls based on gender"
      "LGBTQ+ people": "Gay, lesbian, bisexual, transgender, or other LGBTQ+ individuals"
      "Immigrants": "Immigrants, refugees, or non-citizens"
      "Other group": "Another identifiable group not listed"
      "No specific group": "No clear target group identified"

  # Step 4: Confidence
  - annotation_type: likert
    name: confidence
    description: "How confident are you in your annotation?"
    min_value: 1
    max_value: 5
    labels:
      1: "Very uncertain"
      2: "Somewhat uncertain"
      3: "Moderately confident"
      4: "Confident"
      5: "Very confident"

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