gec-datasets 0.1.0

A tools to easily load Grammatical Error Correction datasets.

gec-datasets

This library is to handle datasets of Grammatical Error Correction.

Usage

from gec_datasets import GECDatasets
gec = GECDatasets(
    base_path='datasets/'
)
conll14 = gec.load('conll14')

assert conll14.srcs is not None
assert conll14.refs is not None
# The number of sentences is 1312.
assert len(conll14.srcs) == 1312
# CoNLL-2014 contains two official references.
assert len(conll14.refs) == 2
# Each reference also contains 1312 sentences.
assert len(conll14.refs[0]) == 1312
assert len(conll14.refs[1]) == 1312

The dataset is stored under base_path=.
The first time it is downloaded automatically, and thereafter it is loaded from the saved files.

datasets/
├── conll14
│   ├── ref0.txt
│   ├── ref1.txt
│   └── src.txt
├── bea19-dev
│   ├── ref0.txt
│   ├── src.txt
├── bea19-test
│   └── src.txt
...

Supported datasets

Public datasets

ID .load(ID)	Description
'conll13'	CoNLL-2013 test set [Ng+ 2013].
'conll14'	CoNLL-2014 test set [Ng+ 2014].
'jfleg-test'	JFLEG test set [Napoles+ 2017].
'jfleg-dev'	JFLEG development set [Napoles+ 2017].
'fce-test'	FCE test set [Yannakoudakis+ 2011]
'fce-dev'	FCE development set [Yannakoudakis+ 2011].
'fce-train'	FCE training set [Yannakoudakis+ 2011].
'cweb-g-test'	CWEB-G test set [Flachs+ 2020].
'cweb-g-dev'	CWEB-G development set [Flachs+ 2020].
'cweb-s-test'	CWEB-S test set [Flachs+ 2020].
'cweb-s-dev'	CWEB-S development set [Flachs+ 2020].
'bea19-test'	BEA-2019 shared task test set [Bryant+ 2019].
'bea19-dev'	BEA-2019 shared task development set [Bryant+ 2019]. It contains only source sentences.
'wi-locness-train'	W&I+LOCNESS training set [Yannakoudakis+ 2018].

Non-public datasets

ID .load(ID)	Description
'nucle-train'	NUCLE training set. [Dahlmeier+ 2013]
'lang8-train'	Lang-8 training set. [Mizumoto+ 2012] [Tajiri+ 2012]

nucle-train

• Request data from HERE.
• You will receive an email with release3.3.tar.bz2 attached.
• mkdir <base_path>/nucle-train/ and put the data as <base_path>/nucle-train/release3.3.tar.bz2.
• You can now use the data with .load("nucle-train"). The data will be extracted automatically.

lang8-train

• Request data from HERE.
• You will receive an email titled "[NAIST Lang-8 Corpus of Learner English for the 14th BEA Shared Task]".
• mkdir <base_path>/lang8-train/ and put the data as <base_path>/lang8-train/lang8.bea19.tar.gz.
• You can now use the data with .load("lang8-train"). The data will be extracted automatically.