KorSTS
KorSTS is the data created and released by KakaoBrain. Data specification is as follows:
- author: KakaoBrain
- repository: https://github.com/kakaobrain/KorNLUDatasets
- references: Ham, J., Choe, Y. J., Park, K., Choi, I., & Soh, H. (2020). KorNLI and KorSTS: New Benchmark Datasets for Korean Natural Language Understanding. arXiv preprint arXiv:2004.03289.
- size:
- train: 5,749 examples
- dev: 1,500 examples
- test: 1,379 examples
Data structure is as:
| Attributes | Property |
|---|---|
| text | Sentence |
| pair | Sentence that makes up a pair with the text |
| label | Relation between the text and pair |
| Others | Additional information regarding the data |
1. In Python
Execute Python console, download the corpus, and read it.
Downloading the corpus
You can download KorSTS in the local by the following procedure.
from Korpora import Korpora
Korpora.fetch("korsts")
First, download the corpus to Korpora, a directory under the user’s local computer root (~/Korpora).If you want to download it in other path, please assign root_dir=custom_path when you execute fetch function.
If you assign force_download=True when you execute the fetch function, the corpus is downloaded again regardless of its presence in the local. The default is False.
Reading the corpus
You can read KorSTS in Python console with the following scheme. If the corpus is not in the local, the downloading is accompanied.
from Korpora import Korpora
corpus = Korpora.load("korsts")
You can read KorSTS as below; the result is the same as the above operation.
from Korpora import KorSTSKorpus
corpus = KorSTSKorpus()
Execute one of the above, and the copus is assigned to the variable corpus. train denotes the train data of KorSTS, and you can check the first instance as:
>>> corpus.train[0]
KorSTSExample(text='비행기가 이륙하고 있다.', pair='비행기가 이륙하고 있다.', label=5.0, genre='main-captions', filename='MSRvid', year='2012test')
>>> corpus.train[0].text
비행기가 이륙하고 있다.
>>> corpus.train[0].pair
비행기가 이륙하고 있다.
>>> corpus.train[0].label
5.0
dev and test denote dev and test data of KorSTS, respectively, and you can check the first instance as:
>>> corpus.dev[0]
# KorSTSExample(text='안전모를 가진 한 남자가 춤을 추고 있다.', pair='안전모를 쓴 한 남자가 춤을 추고 있다.', label=5.0, genre='main-captions', filename='MSRvid', year='2012test')
>>> corpus.test[0]
# KorSTSExample(text='한 소녀가 머리를 스타일링하고 있다.', pair='한 소녀가 머리를 빗고 있다.', label=2.5, genre='main-captions', filename='MSRvid', year='2012test')
The method get_all_texts lets you check all the texts (Sentence) in KorSTS.
>>> corpus.get_all_texts()
['비행기가 이륙하고 있다.', '한 남자가 큰 플루트를 연주하고 있다.', ... ]
The method get_all_pairs lets you check all the pairs (Sentence that makes up a pair with the text) in KorSTS.
>>> corpus.get_all_pairs()
[SentencePair(text='비행기가 이륙하고 있다.', pair='비행기가 이륙하고 있다.'), ... ]
The method get_all_labels lets you check all the labels (Relation between the text and pair) in KorSTS.
>>> corpus.get_all_labels()
['5.000', '3.800', ... ]
2. In terminal
You can download the corpus without executing Python console. The command is as below.
korpora fetch --corpus korsts
First, download the corpus to Korpora, a directory under the user’s local computer root (~/Korpora).If you want to download it in other path, please assign --root_dir custom_path when you execute fetch function in the terminal.
If you assign --force_download when you execute fetch function in the terminal, the corpus is downloaded again regardless of its presence in the local.