README.md

metadata

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
• Authors: Jina Song, Eunbi Cho, Sanghoun Song
• GitHub😺
• Kontrast-YesNo_320sentences
• 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 (아니오).
   • 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 아니오).

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 = 전혀 수용 불가능함 (Totally unacceptable)
     • 2 = 수용 불가능함 (Unacceptable)
     • 3 = 보통임 (Neutral)
     • 4 = 수용 가능함 (Acceptable)
     • 5 = 매우 수용 가능함 (Very acceptable)
   • Columns:
     • id: Unique identifier for each sentence.
     • sentence: The sentence being evaluated.
     • gold_label: Acceptability judgment (정문 or 비문).

Example Data

Forced Choice Task

ID	Acceptable Sentence (A)	Less Acceptable Sentence (B)	Correct Answer
1	서울은 한국의 수도이다.	서울은 한국의 수도뿐이다.	A
2	철수가 어제 준 것은 영희에게 책이야.	철수가 어제 영희에게 준 것은 책이야.	B

Yes/No Task

ID	Sentence	Judgment
1	철수가 어제 영희에게 준 것은 책이야.	예
2	빈번히 일어나는 유괴 사건이 우리를 슬펐게 한다.	아니오

Likert Scale Task

ID	Sentence	Judgment
1	서울은 한국의 수도이다.	정문
2	영이가 예쁘지 않고 있다.	비문

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