|
--- |
|
author: Juan Alberto López Cavallotti |
|
date: Jan 6, 2023 |
|
license: apache-2.0 |
|
task_categories: |
|
- translation |
|
language: |
|
- en |
|
- es |
|
- fr |
|
- de |
|
tags: |
|
- grammar |
|
- gec |
|
- multi language |
|
- language detection |
|
pretty_name: Multi Lingual Grammar Error Correction Dataset |
|
size_categories: |
|
- 100K<n<1M |
|
--- |
|
|
|
# Dataset Card for Multilingual Grammar Error Correction |
|
|
|
## Dataset Description |
|
|
|
- **Homepage:** https://juancavallotti.com |
|
- **Paper:** https://blog.juancavallotti.com/2023/01/06/training-a-multi-language-grammar-error-correction-system/ |
|
- **Point of Contact:** Juan Alberto López Cavallotti |
|
|
|
### Dataset Summary |
|
|
|
This dataset can be used to train a transformer model (we used T5) to correct grammar errors in simple sentences written in English, Spanish, French, or German. |
|
This dataset was developed as a component for the [Squidigies](https://squidgies.app/) platform. |
|
|
|
### Supported Tasks and Leaderboards |
|
|
|
* **Grammar Error Correction:** By appending the prefix *fix grammar:* to the prrompt. |
|
* **Language Detection:** By appending the prefix: *language:* to the prompt. |
|
|
|
### Languages |
|
|
|
* English |
|
* Spanish |
|
* French |
|
* German |
|
|
|
## Dataset Structure |
|
|
|
### Data Instances |
|
|
|
The dataset contains the following instances for each language: |
|
* German 32282 sentences. |
|
* English 51393 sentences. |
|
* Spanish 67672 sentences. |
|
* French 67157 sentences. |
|
|
|
### Data Fields |
|
|
|
* `lang`: The language of the sentence |
|
* `sentence`: The original sentence. |
|
* `modified`: The corrupted sentence. |
|
* `transformation`: The primary transformation used by the synthetic data generator. |
|
* `sec_transformation`: The secondary transformation (if any) used by the synthetic data generator. |
|
|
|
### Data Splits |
|
|
|
* `train`: There isn't a specific split defined. I recommend using 1k sentences sampled randomly from each language, combined with the SacreBleu metric. |
|
|
|
## Dataset Creation |
|
|
|
### Curation Rationale |
|
|
|
This dataset was generated synthetically through code with the help of information of common grammar errors harvested throughout the internet. |
|
|
|
### Source Data |
|
|
|
#### Initial Data Collection and Normalization |
|
|
|
The source grammatical sentences come from various open-source datasets, such as Tatoeba. |
|
|
|
#### Who are the source language producers? |
|
|
|
* Juan Alberto López Cavallotti |
|
|
|
### Annotations |
|
|
|
#### Annotation process |
|
|
|
The annotation is automatic and produced by the generation script. |
|
|
|
#### Who are the annotators? |
|
|
|
* Data generation script by Juan Alberto López Cavallotti |
|
|
|
### Other Known Limitations |
|
|
|
The dataset doesn't cover all the possible grammar errors but serves as a starting point that generates fair results. |
|
|
|
## Additional Information |
|
|
|
### Dataset Curators |
|
|
|
* Juan Alberto López Cavallotti |
|
|
|
### Licensing Information |
|
|
|
This dataset is distributed under the [Apache 2 License](https://www.apache.org/licenses/LICENSE-2.0) |
|
|
|
### Citation Information |
|
|
|
Please mention this original dataset and the author **Juan Alberto López Cavallotti** |
|
|
|
### Contributions |
|
|
|
* Juan Alberto López Cavallotti |
