Each MultiGEC subcorpus is composed of a train, a development and a test set, each of which consists of 2+ essay-aligned files, one containing original learner essays and one or more containing reference (i.e. corrected/normalized) texts.

Naming conventions

Data files are named according to the following convention:

langcode-corpus-orig|refn-train|dev|test.md

where

  • langcode is the two-letter ISO 639 code for the language
  • corpus is the name of the subcorpus, lowercased
  • orig indicates that the file contains original essays, whereas ref1, ref2, refn indicate the n-th reference file
  • test, dev and train indicate the corresponding dataset splits

Example: cs-natwebinf-ref2-dev.md is the file containing the second reference set for the development split of the NatWebInf Czech subcorpus.

File format

Internally, each file follows this simple Markdown-based format:

### essay_id = 1
Full text of the first essay/reference.

Whitespace, including newline characters, is preserved, but for the sake of readability TWO consecutive newline characters spearate subsequent essays.

### essay_id = 2
Full text of the second essay/reference.

...

orig and refn files are aligned at the essay level in the sense that reference corrections with essay_id = Y are relative to the essay with essay_id = Y in the corresponding orig file. Note, however, that not all refn files contain corrections for all essays in their orig counterpart (that is, some subcorpora only have a second reference for some of the essays).

Metadata

Each subcorpus is assicuated with a README file and YAML file that summarizes basic metadata in a machine-readable format:

target_language: two-letter ISO language code of the texts in the corpus, e.g. "sv" for Swedish
source_corpus: name of the source corpus
learner_type: L1|L2|mixed # whether the authors of the essays are first-language speakers, second language learners or both/unknown (e.g if the corpus is crowdsources)
short_description: short description of the corpus
links: # optional
  - link to paper
  - link to data sheet or similar 
  - ...
contact: maintainer@institution.xx # contact information for the main data provider
availability: open|restricted # "open" if the subcorpus is free to use outside of the shared task, restricted otherwise. This refers to the MultiGEC subcorpus, not necessarily to the source corpus
license: 
  name: name of the license
  url: link to the full text of the license
sentence_aligned: false|true # true if the source corpus is sentence-aligned, false otherwise
original_essays: # stats in reference to the *-orig-*.md files
  n_essays: 
    total: a+b+c
    train: a
    dev: b
    test: c
reference_essays_1: # stats in reference to the *-refX-*.md files
  correction_style: minimal|fluency
  n_essays: 
    total: a+b+c
    train: a
    dev: b
    test: c
# reference_essays_2, reference_essays_3...