EunB2's picture
Update README.md
7f46c4a verified
---
dataset_info:
features:
- name: id
dtype: int64
- name: sentence
dtype: string
- name: gold_label
dtype: string
splits:
- name: test
num_bytes: 20371
num_examples: 320
download_size: 10171
dataset_size: 20371
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
task_categories:
- text-classification
- zero-shot-classification
language:
- ko
tags:
- syntax
- acceptability
- minimalpairs
size_categories:
- n<1K
---
# Kontrast Dataset
* Paper: [Evaluating GPT’s Ability to Understand Syntactic Minimal Pairs in Korean](https://doi.org/10.29403/LI.28.3.5)
* Authors: [Jina Song](https://english.hongik.ac.kr/english/0201.do?mode=view&deptCd=AAD140&S1=2024&S2=10024), [Eunbi Cho](https://github.com/EunB2), [Sanghoun Song](http://corpus.mireene.com/)
* [GitHub](https://github.com/EunB2/Kontrast)😺
* [Kontrast-YesNo_320sentences](https://huggingface.co/datasets/EunB2/Kontrast-YesNo_320sentences)
* [Kontrast-ForcedChoice_160pairs](https://huggingface.co/datasets/EunB2/Kontrast-ForcedChoice_160pairs)
This dataset, **Kontrast**, contains Korean syntactic minimal pairs used to evaluate the syntactic competence of large language models (LLMs), including GPT-3.5, GPT-4, and GPT-4o.
## Main Concept
The dataset consists of **syntactic minimal pairs**, where each pair includes:
* **An acceptable sentence**
* **A less acceptable sentence** (due to a syntactic violation)
These pairs help assess whether language models align with **native Korean speaker judgments** regarding syntactic acceptability.
## Data Description
This dataset consists of three subsets based on different experimental tasks:
1. **Forced Choice Task (`ForcedChoice_160pairs.xlsx`)**
- **160 sentence pairs** where one sentence is grammatically more acceptable than the other.
- The model is asked to choose the more acceptable sentence.
- **Columns:**
- `id`: Unique identifier for the sentence pair.
- `sentence_A`: The more acceptable sentence.
- `sentence_B`: The less acceptable sentence.
- `gold_label`: Correct answer (either `A` or `B`).
2. **Yes/No Task (`YesNo_320sentences.xlsx`)**
- **320 individual sentences** labeled as acceptable (`예`) or unacceptable (`μ•„λ‹ˆμ˜€`).
- The model is asked to determine whether each sentence is acceptable.
- **Columns:**
- `id`: Unique identifier for each sentence.
- `sentence`: The sentence being evaluated.
- `gold_label`: Acceptability judgment (`예` or `μ•„λ‹ˆμ˜€`).
3. **Likert Scale Task (`LikertScale_320sentences.xlsx`)**
- **320 individual sentences**, each rated based on **acceptability judgments** by human annotators.
- The model assigns a score between **1 and 5**, where:
- **1 = μ „ν˜€ 수용 λΆˆκ°€λŠ₯함 (Totally unacceptable)**
- **2 = 수용 λΆˆκ°€λŠ₯함 (Unacceptable)**
- **3 = λ³΄ν†΅μž„ (Neutral)**
- **4 = 수용 κ°€λŠ₯함 (Acceptable)**
- **5 = 맀우 수용 κ°€λŠ₯함 (Very acceptable)**
- **Columns:**
- `id`: Unique identifier for each sentence.
- `sentence`: The sentence being evaluated.
- `gold_label`: Acceptability judgment (`μ •λ¬Έ` or `λΉ„λ¬Έ`).
### Example Data
#### **Forced Choice Task**
| ID | Acceptable Sentence (A) | Less Acceptable Sentence (B) | Correct Answer |
|----|-------------------------|-----------------------------|----------------|
| 1 | μ„œμšΈμ€ ν•œκ΅­μ˜ μˆ˜λ„μ΄λ‹€. | μ„œμšΈμ€ ν•œκ΅­μ˜ μˆ˜λ„λΏμ΄λ‹€. | A |
| 2 | μ² μˆ˜κ°€ μ–΄μ œ μ€€ 것은 μ˜ν¬μ—κ²Œ 책이야. | μ² μˆ˜κ°€ μ–΄μ œ μ˜ν¬μ—κ²Œ μ€€ 것은 책이야. | B |
#### **Yes/No Task**
| ID | Sentence | Judgment |
|----|---------|----------|
| 1 | μ² μˆ˜κ°€ μ–΄μ œ μ˜ν¬μ—κ²Œ μ€€ 것은 책이야. | 예 |
| 2 | 빈번히 μΌμ–΄λ‚˜λŠ” 유괴 사건이 우리λ₯Ό 슬펐게 ν•œλ‹€. | μ•„λ‹ˆμ˜€ |
#### **Likert Scale Task**
| ID | Sentence | Judgment |
|----|---------|----------|
| 1 | μ„œμšΈμ€ ν•œκ΅­μ˜ μˆ˜λ„μ΄λ‹€. | μ •λ¬Έ |
| 2 | μ˜μ΄κ°€ μ˜ˆμ˜μ§€ μ•Šκ³  μžˆλ‹€. | λΉ„λ¬Έ |
## Citation
```
@article{song2024evaluating,
author = {Jina Song and Eunbi Cho and Sanghoun Song},
title = {Evaluating GPT’s Ability to Understand Syntactic Minimal Pairs in Korean},
journal = {Language and Information},
volume = {28},
number = {3},
pages = {83-109},
year = {2024},
publisher = {The Korean Society for Language and Information}
}
```
## License
TBD