HateXplain - Explainable Hate Speech Detection

# HateXplain - Explainable Hate Speech Detection
# Based on Mathew et al., AAAI 2021
# Paper: https://ojs.aaai.org/index.php/AAAI/article/view/17745
# Dataset: https://huggingface.co/datasets/hatexplain
#
# Three annotation tasks:
# 1. Classification: hate speech, offensive, or normal
# 2. Target community: which group is targeted (if hate/offensive)
# 3. Rationale spans: which words justify the classification
#
# Guidelines:
# - Hate speech: attacks or demeans a group based on identity
# - Offensive: rude/disrespectful but not targeting identity groups
# - Normal: neither hateful nor offensive
# - Rationale: highlight words that justify your classification (avg 5.5 tokens)

port: 8000
server_name: localhost
task_name: "HateXplain: Explainable Hate Speech Detection"

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

output_file: annotations.json

annotation_schemes:
  # Task 1: Classification
  - annotation_type: radio
    name: classification
    description: "Classify this text as hate speech, offensive, or normal"
    labels:
      - Hate Speech
      - Offensive
      - Normal
    keyboard_shortcuts:
      "Hate Speech": "h"
      "Offensive": "o"
      "Normal": "n"
    tooltips:
      "Hate Speech": "Content that attacks or demeans a group based on identity attributes (race, religion, gender, etc.)"
      "Offensive": "Rude, disrespectful, or profane content that does NOT target identity groups"
      "Normal": "Content that is neither hateful nor offensive"

  # Task 2: Target community (only for hate/offensive)
  - annotation_type: multiselect
    name: target_community
    description: "If hate/offensive, select the targeted community/communities"
    labels:
      - African
      - Arab
      - Asian
      - Caucasian
      - Hispanic
      - Jewish
      - LGBTQ
      - Islam
      - Women
      - Refugee
      - Other
      - None/Not Applicable
    tooltips:
      African: "People of African descent"
      Arab: "People of Arab descent or from Arab countries"
      Asian: "People of Asian descent"
      Caucasian: "People of European/white descent"
      Hispanic: "People of Hispanic/Latino descent"
      Jewish: "Jewish people (ethnic or religious)"
      LGBTQ: "Lesbian, gay, bisexual, transgender, queer individuals"
      Islam: "Muslims or Islamic religion"
      Women: "Women or girls"
      Refugee: "Refugees, immigrants, or asylum seekers"
      Other: "Other identity group not listed"
      "None/Not Applicable": "No specific group targeted (for Normal texts)"

  # Task 3: Rationale span annotation
  - annotation_type: span
    name: rationale
    description: "Highlight the words/phrases that justify your classification decision"
    labels:
      - Rationale
    label_colors:
      Rationale: "#ef4444"
    tooltips:
      Rationale: "Words or phrases that are the reason for classifying as hate/offensive (avg ~5.5 tokens per post)"
    allow_overlapping: false

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