Dynamic Hate Speech Detection

# Dynamic Hate Speech Detection
# Based on Vidgen et al., ACL 2021
# Paper: https://aclanthology.org/2021.acl-long.132/
# Dataset: https://github.com/bvidgen/Dynamically-Generated-Hate-Speech-Dataset
#
# This task uses a hierarchical annotation scheme:
# 1. Binary classification: Hate vs Not Hate
# 2. Hate type (if hateful): 5 categories
# 3. Target group identification
#
# Hate Type Definitions:
# - Animosity: Expression of negative feelings, hostility, or opposition
# - Derogation: Insulting, demeaning, or belittling language
# - Dehumanization: Comparing groups to animals, diseases, or subhuman entities
# - Threatening: Direct or implicit threats of violence or harm
# - Support for Hateful Entities: Praising hateful groups, symbols, or ideologies
#
# Annotation Guidelines:
# 1. Consider the overall message and intent
# 2. Slurs alone may not be hateful (reclaimed language, discussion)
# 3. Criticism of ideas/behaviors differs from attacks on identity groups
# 4. Context matters - sarcasm, quotes, and counter-speech should be considered
# 5. When uncertain, consider how a member of the target group would perceive it

annotation_task_name: "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: Binary hate classification
  - annotation_type: radio
    name: is_hateful
    description: "Is this content hateful?"
    labels:
      - "Hate"
      - "Not Hate"
    tooltips:
      "Hate": "Content that attacks, demeans, or threatens individuals or groups based on protected characteristics"
      "Not Hate": "Content that is not hateful, including legitimate criticism, counter-speech, or neutral discussion"

  # Step 2: Hate type classification (if hateful)
  - annotation_type: radio
    name: hate_type
    description: "What type of hate is expressed? (Select the primary type)"
    labels:
      - "Animosity"
      - "Derogation"
      - "Dehumanization"
      - "Threatening"
      - "Support for Hateful Entities"
    tooltips:
      "Animosity": "Expression of negative feelings, hostility, antipathy, or opposition toward a group"
      "Derogation": "Insulting, demeaning, belittling, or degrading language about a group"
      "Dehumanization": "Comparing a group to animals, vermin, diseases, or subhuman entities"
      "Threatening": "Direct or implicit threats of violence, harm, or calls for exclusion/extermination"
      "Support for Hateful Entities": "Praising, supporting, or glorifying hateful groups, symbols, figures, or ideologies"

  # Step 3: Target group identification
  - annotation_type: multiselect
    name: target_groups
    description: "Which groups are targeted? (Select all that apply)"
    labels:
      - "Black people"
      - "Asian people"
      - "Hispanic/Latino people"
      - "Arab/Middle Eastern people"
      - "Indigenous people"
      - "Jewish people"
      - "Muslims"
      - "Women"
      - "LGBTQ+ people"
      - "Trans people"
      - "People with disabilities"
      - "Immigrants/Refugees"
      - "Other group"
    label_colors:
      "Black people": "#3b82f6"
      "Asian people": "#06b6d4"
      "Hispanic/Latino people": "#8b5cf6"
      "Arab/Middle Eastern people": "#f59e0b"
      "Indigenous people": "#22c55e"
      "Jewish people": "#6366f1"
      "Muslims": "#14b8a6"
      "Women": "#ec4899"
      "LGBTQ+ people": "#a855f7"
      "Trans people": "#d946ef"
      "People with disabilities": "#f97316"
      "Immigrants/Refugees": "#0ea5e9"
      "Other group": "#9ca3af"
    min_selections: 0
    max_selections: 13

  # 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