WikiANN (PAN-X): Multilingual Named Entity Recognition

# WikiANN / PAN-X - Multilingual Named Entity Recognition
# Based on Pan et al., ACL 2017 (original) and Rahimi et al., ACL 2019 (balanced splits)
# Papers: https://aclanthology.org/P17-1178/  |  https://aclanthology.org/P19-1015/
# Dataset: https://huggingface.co/datasets/unimelb-nlp/wikiann
#
# IMPORTANT — how WikiANN was made: WikiANN (a.k.a. PAN-X) is a SILVER-standard,
# automatically generated resource. Pan et al. (2017) projected entity labels
# from Wikipedia anchor-link markup and knowledge-base types (DBpedia, YAGO,
# Freebase) across 282 languages. There were no human annotators and no
# annotation guideline — so the labels carry known noise. The widely used
# distribution is Rahimi et al.'s (2019) balanced re-split of 176 languages,
# which uses three tags — PER, ORG, LOC — in IOB2 format.
#
# This Potato config lets you produce GOLD labels in the same PER/ORG/LOC scheme
# — either annotating fresh Wikipedia sentences in any language, or correcting
# silver WikiANN labels. It is not a reconstruction of an original human protocol
# (there wasn't one); it is the human-in-the-loop task WikiANN's noise motivates.
#
# Tag set (Rahimi et al. 2019 balanced splits):
# - PER (Person): names of people
# - ORG (Organization): companies, institutions, agencies, teams
# - LOC (Location): countries, cities, regions, and other place names
#   (the original paper's GPE type is folded into LOC in the standard splits)

annotation_task_name: "WikiANN - Multilingual NER"
task_dir: "."

data_files:
  - sample-data.json

item_properties:
  id_key: "id"
  text_key: "text"

output_annotation_dir: "annotation_output/"
output_annotation_format: "json"

port: 8000
server_name: localhost

annotation_schemes:
  # Highlight entity spans and label each with its type
  - annotation_type: span
    name: entity_spans
    description: "Highlight every named entity and label it PER, ORG, or LOC"
    labels:
      - "PER (Person)"
      - "ORG (Organization)"
      - "LOC (Location)"
    label_colors:
      "PER (Person)": "#3b82f6"
      "ORG (Organization)": "#22c55e"
      "LOC (Location)": "#ef4444"
    keyboard_shortcuts:
      "PER (Person)": "1"
      "ORG (Organization)": "2"
      "LOC (Location)": "3"

  # Whether the auto-projected silver labels (if any) were correct
  - annotation_type: radio
    name: silver_label_check
    description: "If this sentence came with silver WikiANN labels, were they correct?"
    labels:
      - "Correct"
      - "Needs fixing"
      - "No silver labels"
    keyboard_shortcuts:
      "Correct": "4"
      "Needs fixing": "5"
      "No silver labels": "6"
    tooltips:
      "Correct": "The pre-existing WikiANN labels matched your annotation"
      "Needs fixing": "The pre-existing WikiANN labels were wrong or incomplete"
      "No silver labels": "This sentence was annotated from scratch"

annotation_instructions: |
  You will see a short sentence, typically from Wikipedia, in one of many
  languages. Using the WikiANN / PAN-X tag set (PER, ORG, LOC):

  1. Highlight every named-entity mention and label it with the correct type.
     Select the full span (e.g. "United Nations", not just "United").
  2. A place name inside an organization name is part of the ORG span (e.g. in
     "University of Nairobi", the whole phrase is ORG).
  3. If the sentence arrived with silver WikiANN labels, record whether those
     labels were correct or needed fixing.

  Note: WikiANN labels are generated automatically from Wikipedia links, so they
  are noisy. Your job is to produce clean, gold-standard PER/ORG/LOC labels.

html_layout: |
  <div style="padding: 15px; max-width: 800px; margin: auto;">
    <div style="background: #eef2ff; border: 1px solid #c7d2fe; border-radius: 8px; padding: 8px 12px; margin-bottom: 12px; font-size: 13px; color: #4338ca;">
      Language: {{lang}}
    </div>
    <div style="background: #f0f9ff; border: 1px solid #bae6fd; border-radius: 8px; padding: 16px;">
      <p style="font-size: 18px; line-height: 1.8; margin: 0;">{{text}}</p>
    </div>
  </div>

allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 2
allow_skip: true
skip_reason_required: false