CLAMS - Cross-Linguistic Assessment of Models on Syntax

Dataset Description

CLAMS (Cross-Linguistic Assessment of Models on Syntax) is a dataset of syntactic minimal pairs for English, French, German, Hebrew, and Russian. The dataset contains grammaticality judgment pairs (good/bad sentence pairs) for various syntactic phenomena, designed to evaluate language models' syntactic knowledge across different languages.

Dataset Structure

The dataset contains minimal pairs for 5 languages, each organized as separate subsets:

• en (English): Syntactic minimal pairs for English
• fr (French): Syntactic minimal pairs for French
• de (German): Syntactic minimal pairs for German
• he (Hebrew): Syntactic minimal pairs for Hebrew
• ru (Russian): Syntactic minimal pairs for Russian

Fields

• sentence_good: The grammatical sentence
• sentence_bad: The ungrammatical sentence
• test_case: The specific syntactic test case name (filename without extension)
• pair_id: Unique identifier for the minimal pair within the test case
• good_line: Original line number of the good sentence
• bad_line: Original line number of the bad sentence

Usage

from datasets import load_dataset

# Load specific language
dataset = load_dataset("juletxara/clams", "en")

# Or load all languages
dataset = load_dataset("juletxara/clams")

# Access specific language data
english_data = dataset['en']

# Example usage
for example in english_data:
    print(f"Good: {example['sentence_good']}")
    print(f"Bad: {example['sentence_bad']}")
    print(f"Test case: {example['test_case']}")
    print()

Data Source

This dataset is derived from the CLAMS project: https://github.com/aaronmueller/clams

Citation

Please cite the original CLAMS paper when using this dataset:

@inproceedings{mueller-etal-2020-cross,
    title = "Cross-Linguistic Syntactic Evaluation of Word Prediction Models",
    author = "Mueller, Aaron and Linzen, Tal and Dunagan, Dustin and Bau, David and Belinkov, Yonatan",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    year = "2020",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.acl-main.490",
    pages = "5523--5539"
}

License

This dataset follows the original CLAMS project's Apache License, Version 2.0.