This subcorpus currently consists of 263 texts and 3,790 sentences retrieved from the Estonian Interlanguage Corpus compiled at the Tallinn University. The texts include narrative/descriptive and argumentative writings as well as informal and formal letters representing various proficiency levels. EstGEC-L2 material has been divided into a test and development set that can be used for evaluating and improving Estonian automated correction tools. The test set comprises 2,029 and the dev set 1,761 sentences, distributed between the proficiency levels as follows:

• A2 – 937 (495 in test set);
• B1 – 963 (504 in test set);
• B2 – 1,091 (534 in test set);
• C1 – 796 (495 in test set).

Previously, the texts had been manually error-tagged in the CoNLL-U format, indicating the error type, scope, and correction in the field for miscellaneous token attributes. The annotation has been converted to the M2 format (the conversion script can be found here) using an adapted version of the ERRANT tagset. Whereas the previous format was limited to one error annotation per sentence, up to two new annotation versions have been added. Considering the two-phase annotation, each text has been reviewed by at least three annotators.

There are 12 main and 18 combined error types in the error classification (see tables 1 and 2). The prefix indicates whether a word, phrase or punctuation mark should be replaced ('R:'), is missing ('M:') or unnecessary ('U:'). In our tagset, we do not distinguish the part-of-speech (POS) of the replaced, added or deleted word. For example, all word choice errors are indicated by the tag 'R:LEX'. This has helped to reduce the complexity of the error categorization, while allowing us to classify all errors and avoid the 'OTHER' tag. There are numerous edit and POS combinations, since the edit types often overlap (e.g., spelling errors co-occur with inflection and word choice errors) and the POS of the original word and its replacement can differ.

Another important difference to the English M2 annotation is that we allow overlapping error scope if a token-level error occurs within a word order error, e.g., one of the words contains a spelling error. Therefore, it is possible to detect token-level corrections even if word order has not been edited.

Furthermore, orthography errors have been divided into capitalization and whitespace errors. Inflection errors are marked as nominal (noun, adjective, pronoun and numeral) or verb form errors without a further distinction, i.e., these include case, number, agreement, tense, mood and other errors in the choice of inflected form.

Table 1. Main error types

Table 2. Combined error types

• The dataset has been used to evaluate the GEC toolkit developed in collaboration by the language technology groups of the University of Tartu and the Tallinn University. The L1 subset of the EstGEC corpus is being annotated at the University of Tartu.
• The M2 Scorer adapted for EstGEC can be found here.
• Conference presentations:
  • 8th Estonian Digital Humanities Conference, October 5-7, 2022, Tallinn
  • 20th Annual Conference of Applied Linguistics, April 27-28, 2023, Tallinn (in Estonian)