---
license: gpl-3.0
task_categories:
- translation
language:
- et
tags:
- GEC
- Grammatical Error Correction
pretty_name: Erroneous Estonian Sentences from Non-Native Speakers by Tallinn University
size_categories:
- 1K<n<10K
configs:
- config_name: Hugging Face Preview
  data_files: Hugging_Face.csv
  sep: '~'
---

# Tallinn-L2 Corpus

## Dataset Summary

The **Tallinn-L2 corpus** is a significant dataset developed by the Language Technology Research Group at Tallinn University. This corpus is a valuable resource for Grammatical Error Correction (*GEC*) research, particularly in the Estonian language. The dataset consists of 3,790 sentences annotated in the Max-Match (*M2*) format, highlighting the type and location of errors made by learners of Estonian.

### Features

- **Total Sentences:** 2,029 sentences (*test subset from source*)
- **Learner Levels:** A2, B1, B2, and C1 levels according to the Common European Framework of Reference for Languages (*[CEFR](https://www.coe.int/en/web/common-european-framework-reference-languages/level-descriptions)*)
- **Annotations:** Sentences are annotated using the Estonian Max-Match (*M2*) format, which identifies both the type of error and its specific location within the sentence.
- **Source:** Derived from the Estonian Interlanguage Corpus, providing a diverse and balanced dataset.

### Dataset Composition

The corpus features sentences from learners across a range of proficiency levels:

- **A2 Level:** Beginner learners
- **B1 Level:** Intermediate learners
- **B2 Level:** Upper-intermediate learners
- **C1 Level:** Advanced learners

## Supported Tasks and Leaderboards

The Tallinn-L2 corpus is useful for several tasks, including:

- **Grammatical Error Correction (*GEC*):** Ideal for training and evaluating models designed to correct grammatical errors in Estonian.
- **Second Language Acquisition (*SLA*) Research:** Provides insights into the learning behaviors and common errors among learners of Estonian as a second language.

## Languages

The primary language of the corpus is **Estonian**, with sentences generated by learners of Estonian from various native language backgrounds.

## Dataset Structure

### Data Instances

Each instance in the dataset includes:

- **Original Sentence:** The uncorrected sentence as written by the learner.
- **Annotations:** Detailed error annotations in the Estonian Max-Match (*M2*) format, including:
  - Type of error (*e.g., grammatical, lexical*)
  - Location of the error within the sentence

### Data Splits

The dataset is not split into predefined training, validation, or test sets. Users may need to create their own splits depending on the task.

## Dataset Creation

### Curation Rationale

The Tallinn-L2 corpus was created to fill a gap in resources for grammatical error correction in the Estonian language. It serves both linguistic research and the development of Natural Language Processing (*NLP*) tools for Estonian.

### Source Data

- **Initial Data Collection:** Sentences were collected from learners of Estonian at various proficiency levels.
- **Annotations:** Sentences were annotated using the Estonian Max-Match (*M2*) format, identifying specific errors and their locations within the text.

### Licencing Information

Language Technology Research Group's [GitHub release](https://github.com/tlu-dt-nlp/EstGEC-L2-Corpus/blob/main/LICENSE) has a GNU General Public Licence v3.0.

## Citation Information

If you use the Tallinn-L2 corpus in your research, feel free to use the following reference:
```
inproceedings{Tallinn-L2,
  title  = {{E}rror {C}lassification \& {A}nnotation of {L}earner {L}anguage for {D}eveloping {E}stonian {G}rammar {C}orrection},
  author = {Kais Allkivi-Metsoja and Karina Kert and Silvia Maine and Kaisa Norak and Pille Eslon},
  booktitle = {Proceedings of the 8th Estonian Digital Humanities Conference},
  year = {2022},
  date = {October 5-7},
  address = {Tallinn, Estonia}
}
```
## Contributions

Thanks to Kais Allkivi-Metsoja and the Language Technology Research Group at Tallinn University for creating and maintaining this corpus. Contributions to the dataset are welcome, and you can reach out to the dataset maintainers for more information.