SIGMORPHON 2023 Lezgi Interlinear Glossing with ByT5 Pre-Labels

# Lezgi Interlinear Glossing (SIGMORPHON 2023 data format, placeholder items)
#
# Dataset: SIGMORPHON 2023 Shared Task on Interlinear Glossing, Lezgi subset
#   (701 training and 88 development sentences, per the shared-task README)
# Dataset URL: https://github.com/sigmorphon/2023glossingST
# Licence: Lezgi data is CC BY-NC 4.0 (data_v1/Lezgi/LICENSE). Code is Apache-2.0.
#   Because the licence is non-commercial, NO sentence from the shared-task
#   files is reproduced here. Every item in sample-data.json is an artificial placeholder.
# DOI: 10.18653/v1/2023.sigmorphon-1.20
#
# Paper (ACL Anthology BibTeX):
#   @inproceedings{ginn-etal-2023-findings,
#     title = "Findings of the {SIGMORPHON} 2023 Shared Task on Interlinear Glossing",
#     author = "Ginn, Michael and Moeller, Sarah and Palmer, Alexis and Stacey, Anna
#               and Nicolai, Garrett and Hulden, Mans and Silfverberg, Miikka",
#     booktitle = "Proceedings of the 20th {SIGMORPHON} workshop on Computational
#                  Research in Phonetics, Phonology, and Morphology",
#     month = jul, year = "2023", address = "Toronto, Canada",
#     publisher = "Association for Computational Linguistics",
#     url = "https://aclanthology.org/2023.sigmorphon-1.20/",
#     doi = "10.18653/v1/2023.sigmorphon-1.20", pages = "186--201" }
#
# Data format quoted from the shared-task README (open track, track2):
#   \t orthographic representation
#   \m morphological segmentation
#   \g gold standard gloss
#   \l English (or Spanish) translation
# and the task statement: "Participants build systems which generate
# morpheme-level grammatical descriptions of input sentences following the
# Leipzig glossing conventions". The README says the Lezgi data has
# segmentations and English translations but no POS tags, so no extra tier is
# configured. Item fields map as \t = text, \m = morphemes, \g = glosses,
# \l = translation.
#
# THE SAMPLE ITEMS ARE ARTIFICIAL PLACEHOLDERS. The words are STEM1, STEM2 and
# so on, the suffixes are sfx1 to sfx7, and the glosses name each stem lexN and
# each suffix with a standard Leipzig tag (sfx1 = ERG, sfx2 = PFV, sfx3 = PL,
# sfx4 = DAT, sfx5 = GEN, sfx6 = OBL, sfx7 = ABL). Nothing here is Lezgi text,
# and the translations say so. They exist to show the format, the Leipzig
# boundary checks and the training loop. Replace them with sentences you are
# licensed to use. The language key is left unset for that reason.
#
# This design is the Lezgi track with the byt5-gloss training loop. The older
# design text/parsing/sigmorphon-interlinear-glossing covers several languages
# of the same shared task with a three-scheme workflow.
#
# Items 1-4 arrive fully glossed (a review task), items 5-7 have a segmentation
# but no glosses (the open-track "covered" input), items 8-10 are raw text.
#
# Leipzig conventions in use: "-" joins affixes, "=" joins clitics, "." joins
# several category labels of one morpheme, and grammatical categories are in
# capitals (ERG, DAT, PL, PFV). The segmentation and the gloss of a word must
# have the same boundaries in the same order; the page flags a word that does
# not, and flags category labels outside the Leipzig list.
#
# Adaptations:
#   - Items are artificial placeholders, not taken from the shared task (licence above).
#   - One sentence per item; the shared task files hold many sentences each.
#   - Gloss correction, not just generation: some items arrive pre-glossed.
#   - The trainer is Potato's; the shared task systems and scores are not
#     reproduced and no accuracy is claimed.
#
# MODEL TRAINING: the byt5-gloss trainer (pip install 'potato-annotation[train-text]')
# fine-tunes ByT5 to segment words into morphemes and to gloss each morpheme,
# using the words annotators have glossed. Writeback shows its predictions as
# prelabels. Open /admin/training with the admin key below to pick the trainer.
# It does not serve a model. Ten glossed words is the minimum
# (min_instances), far too few for a useful model; the ten sample sentences
# only let the loop run.

annotation_task_name: "Lezgi Interlinear Glossing (placeholder items)"

task_dir: "."
data_files:
  - sample-data.json

item_properties:
  id_key: id
  text_key: text

output_annotation_dir: "annotation_output/"
export_annotation_format: "jsonl"

require_password: false
admin_api_key: "demo-admin-key"

user_config:
  allow_all_users: true
  users: []

annotation_instructions: |
  Each item is a sentence in the SIGMORPHON 2023 Lezgi format. Split each word
  into morphemes in the segmentation row and give one gloss per morpheme in the
  gloss row, following the Leipzig Glossing Rules. Use - between affixes and =
  before a clitic, in both rows. Write grammatical categories in capitals (ERG,
  DAT, PL, PFV), write lexical meanings in lowercase, and join several meanings
  of one morpheme with a dot. The sentences here are artificial placeholders and are
  not taken from the shared-task data.

annotation_schemes:
  - annotation_type: interlinear_gloss
    name: igt
    description: "Segment each word into morphemes and gloss every morpheme"
    instructions: >-
      Put - between affixes and = before a clitic, in both rows. Write
      grammatical categories in capitals and lexical meanings in lowercase.
      Stems in these sample sentences are artificial placeholders.
    source_field: text
    morphemes_field: morphemes
    gloss_field: glosses
    translation_field: translation
    abbreviations: leipzig
    metalanguage: en
    label_requirement:
      required: true

model_training:
  enabled: true
  writeback:
    schemas: ["igt"]
    include_unlabeled: true
    min_confidence: 0.6