Toxic Spans Detection

# Toxic Spans Detection
# Based on SemEval-2021 Task 5 (Pavlopoulos et al., 2021)
# Paper: https://aclanthology.org/2021.semeval-1.6/
# Dataset: https://github.com/ipavlopoulos/toxic_spans
#
# Task: Identify the specific character sequences within comments that
# contribute to toxicity, rather than making binary judgments about
# entire comments.
#
# Guidelines:
# - Mark the exact words/phrases that make the text toxic
# - Focus on language that is abusive, offensive, or harmful
# - Be precise: highlight only the toxic portions, not surrounding context
# - Multiple spans can be marked in a single comment
# - Some comments may have no toxic spans (false positives in toxicity detection)

port: 8000
server_name: localhost
task_name: "Toxic Spans Detection"

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

output_file: annotations.json

annotation_schemes:
  # First: determine if the text contains toxicity
  - annotation_type: radio
    name: contains_toxicity
    description: "Does this text contain any toxic content?"
    labels:
      - "Yes - contains toxic content"
      - "No - not toxic"
    keyboard_shortcuts:
      "Yes - contains toxic content": "y"
      "No - not toxic": "n"
    tooltips:
      "Yes - contains toxic content": "The text contains language that is abusive, offensive, or harmful"
      "No - not toxic": "The text does not contain toxic language (may be critical but not abusive)"

  # Then: highlight the specific toxic spans
  - annotation_type: span
    name: toxic_spans
    description: "Highlight the specific words or phrases that make this text toxic"
    labels:
      - Toxic
    label_colors:
      Toxic: "#ef4444"
    tooltips:
      Toxic: "Words or phrases that are abusive, offensive, threatening, or otherwise harmful"
    allow_overlapping: false

  # Optional: categorize the type of toxicity
  - annotation_type: multiselect
    name: toxicity_type
    description: "What type(s) of toxicity are present? (select all that apply)"
    labels:
      - Insult
      - Profanity
      - Threat
      - Identity Attack
      - Sexual Content
      - Other
    tooltips:
      Insult: "Personal attacks or demeaning language"
      Profanity: "Vulgar or obscene language"
      Threat: "Expressions of intent to harm"
      "Identity Attack": "Attacks based on identity (race, gender, religion, etc.)"
      "Sexual Content": "Sexually explicit or inappropriate content"
      Other: "Other forms of toxic content"

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