multiCHILDES: Multilingual Child-Directed Speech Dataset

Dataset Description

This dataset contains child-directed speech from 19 languages, extracted from the CHILDES corpus. The text has been cleaned and is designed for text generation tasks, particularly in studying early language acquisition.

Dataset Details

• Source: CHILDES corpus
• Languages: 19 languages
• Text Type: Child-directed speech
• Task: Text Generation, Language Modeling
• Data Processing: The dataset has been cleaned to remove inconsistencies and noise.

Intended Use

This dataset is suitable for:

• Training and evaluating language models for early language acquisition
• Studying cross-linguistic differences in child-directed speech
• Supporting research in psycholinguistics and computational linguistics

Usage

To load the dataset using datasets from Hugging Face:

from datasets import load_dataset

dataset = load_dataset("IParraMartin/multiCHILDES")
print(dataset["train"][0])

Data Statistics

• Total Samples: 76529
• Average Utterance Length: 500

Data Collection and Preprocessing

• Source: Extracted from CHILDES (MacWhinney, 1998)
• Preprocessing Steps:
  • Removed extraneous annotations
  • Normalized text
  • Filtered out noisy data

Limitations & Biases

• The dataset is limited to child-directed speech, making it unsuitable for general-purpose NLP tasks.
• The representation of languages depends on CHILDES availability and data quantity, which may introduce selection bias.
• Possible transcription errors or inconsistencies from the original sources.

Citation

If you use this dataset, please cite:

@article{macwhinney1998childes,
  title={The CHILDES system},
  author={MacWhinney, Brian},
  journal={Handbook of child language acquisition},
  pages={457--494},
  year={1998},
  publisher={Brill}
}

Acknowledgments

Special thanks to the CHILDES project for providing high-quality child language data and to the Hugging Face community for enabling open dataset sharing.