ysenarath/Moosa2022Multilingual

Dataset Card for Multilingual Hate Speech Dataset

This dataset card provides information about the Multilingual Hate Speech Dataset, which was originally hosted on Kaggle. The Multilingual Hate Speech Dataset is a modified version of an original multilingual hate speech dataset. In this version, examples from each language have been translated into the other languages present in the dataset, creating a more comprehensive cross-lingual resource.

Dataset Details

Dataset Description

• Curated by: Wajid Hassan Moosa and Najiba
• Shared by: Yasas (for private use only)
• Language(s) (NLP): Arabic, English, Chinese, French, German, Russian, Turkish, Roman Hindi/Urdu, Korean, Italian, Spanish, Portuguese, Indonesian
• License: Open Data Commons Attribution License (ODC-By) v1.0

Dataset Sources

• Repository: https://www.kaggle.com/datasets/wajidhassanmoosa/multilingual-hatespeech-dataset/data

Uses

Direct Use

This dataset is intended for training and evaluating machine learning models for hate speech detection across multiple languages. It can be used for:

• Developing multilingual/cross-lingual hate speech classifiers
• Comparative studies of hate speech patterns across different languages
• Improving content moderation systems for multilingual platforms

Out-of-Scope Use

• This dataset should not be used to make decisions about individuals or to profile users based on their language use.
• It should not be considered a comprehensive representation of hate speech in all contexts or cultures.
• The dataset should not be used for any activity that harms its intended purpose of societal good.

Dataset Structure

The dataset is structured as a collection of records, each containing the following fields:

• text: The content of the message or post (string).
• labels: A binary classification label (integer). 0 indicates non-hate speech, while 1 indicates hate speech.
• lang: The language of the text (string).
• original_lang: The original language of the text (string). This might differ from lang if translation was involved.
• fold: Indicates which subset of the data this sample belongs to (string). One of [train, test, valid].

This dataset was created to enhance an existing multilingual hate speech dataset by providing translations of each example into all other languages present in the dataset. This approach aims to support more robust cross-lingual hate speech detection and analysis.

[More Information Needed]

Bias, Risks, and Limitations

• While the dataset contains original texts in multiple languages, the translations may introduce biases or inaccuracies.
• The effectiveness of hate speech detection models may vary between original texts and translations.
• Cultural and linguistic nuances specific to hate speech in each language may be better preserved in original texts than in translations.
• There may be other biases pertaining to the original dataset.

Recommendations

• Users should consider the origin of each example (original or translation) when training and evaluating models.
• Comparative studies between original texts and their translations could provide insights into the impact of translation on hate speech detection.

Citation

BibTeX:

@misc{wajid_hassan_moosa_najiba_2022,
    title={Multi-lingual HateSpeech Dataset},
    url={https://www.kaggle.com/ds/2260058},
    DOI={10.34740/KAGGLE/DS/2260058},
    publisher={Kaggle},
    author={Wajid Hassan Moosa and Najiba},
    year={2022}
}

APA:

Wajid Hassan Moosa, and Najiba. (2022). Multi-lingual HateSpeech Dataset [Data set]. Kaggle. https://doi.org/10.34740/KAGGLE/DS/2260058