Stance Detection (VAST)

# Stance Detection (VAST-style)
# Based on Allaway & McKeown, EMNLP 2020
# Paper: https://aclanthology.org/2020.emnlp-main.197/
#
# Stance detection determines the author's position toward a target topic.
# This is crucial for understanding public opinion, debate analysis, and
# identifying polarization in social media.
#
# Stance Labels:
# - FAVOR: The text supports or agrees with the topic/claim
# - AGAINST: The text opposes or disagrees with the topic/claim
# - NEUTRAL: The text neither supports nor opposes (discusses without position)
#
# Annotation Guidelines:
# 1. Read the topic/target first, then read the text
# 2. Focus on the AUTHOR's stance, not what they're reporting about
# 3. Implicit stances count - look for sentiment, word choice, framing
# 4. NEUTRAL means genuinely balanced or discussing without taking sides
# 5. Don't confuse "not mentioning" with "neutral" - text must be relevant
# 6. Consider: Would the author agree with "Topic is good/should happen"?
#
# Common Pitfalls:
# - Sarcasm: The literal words may oppose the actual stance
# - Quotations: Distinguish author's stance from quoted opinions
# - Conditional statements: "If X, then Y" may not indicate stance on X

port: 8000
server_name: localhost
task_name: "Stance Detection"

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

output_file: annotations.json

annotation_schemes:
  # Step 1: Stance classification
  - annotation_type: radio
    name: stance
    description: "What is the author's stance toward the given TOPIC?"
    labels:
      - "FAVOR"
      - "AGAINST"
      - "NEUTRAL"
    tooltips:
      "FAVOR": "The author supports, agrees with, or promotes the topic"
      "AGAINST": "The author opposes, disagrees with, or argues against the topic"
      "NEUTRAL": "The author discusses the topic without taking a clear position"

  # Step 2: Stance strength
  - annotation_type: likert
    name: strength
    description: "How strong is the expressed stance?"
    min_value: 1
    max_value: 5
    labels:
      1: "Very weak/implicit"
      2: "Weak"
      3: "Moderate"
      4: "Strong"
      5: "Very strong/explicit"
    tooltips:
      1: "Stance is barely detectable, highly implicit"
      2: "Stance is present but understated"
      3: "Clear stance without strong language"
      4: "Explicit stance with conviction"
      5: "Extremely strong, emphatic stance"

  # Step 3: Confidence
  - annotation_type: likert
    name: confidence
    description: "How confident are you in this 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