The viewer is disabled because this dataset repo requires arbitrary Python code execution. Please consider removing the loading script and relying on automated data support (you can use convert_to_parquet from the datasets library). If this is not possible, please open a discussion for direct help.

Dataset Card for Bernice Pre-train Data

Dataset Summary

Tweet IDs for the 2.5 billion multilingual tweets used to train Bernice, a Twitter encoder. Read the paper here. The tweets are from the public 1% Twitter API stream from January 2016 to December 2021. Twitter-provided language metadata is provided with the tweet ID. The data contains 66 unique languages, as identified by ISO 639 language codes, including und for undefined languages. Tweets need to be re-gathered via the Twitter API. We suggest Hydrator or tweepy.

To load with HuggingFace:

from datasets import load_dataset
dataset = load_dataset("jhu-clsp/bernice-pretrain-data")

for i, row in enumerate(dataset["train"]):
    print(row)
    if i > 10:
        break

If you only want Indic languages, use

dataset = load_dataset("jhu-clsp/bernice-pretrain-data", "indic")

Supported Tasks and Leaderboards

N/A

Languages

65 languages (ISO 639 codes shown below), plus an und (undefined) category. All language identification provided by Twitter API.

en ru ht zh bn ps lt
es bo ur ta sr ckb km
pt it sv ro bg si dv
ja th ca no mr hy lo
ar de el uk ml or ug
in hi fi cy is pa
ko pl cs ne te am
tr nl iw hu gu sd
fr fa da eu kn my
tl et vi sl lv ka

Dataset Structure

Data Instances

Data is provided in gzip'd files organized by year and month of tweet origin. Tweets are one per line, with fields separated by tabs.

Data Fields

  • tweet ID: ID of tweet
  • lang: ISO 639 code of language, provided by Twitter metadata. Accuracy of label is not known.
  • year: Year tweet was created. Year is also provided in the file names.

Data Splits

[More Information Needed]

Dataset Creation

Curation Rationale

Data was gathered to support the training of Bernice, a multilingual pre-trained Twitter encoder.

Source Data

Initial Data Collection and Normalization

Data was gathered via the Twitter API public 1% stream from January 2016 through December 2021. Tweets with less than three non-username or URL space-delimited words were removed. All usernames and URLs were replaced with @USER and HTTPURL, respectively.

Who are the source language producers?

Data was produced by users on Twitter.

Annotations

N/A

Personal and Sensitive Information

As per Twitter guidelines, only tweet IDs and not full tweets are shared. Tweets will only be accessible if user has not removed their account (or been banned), tweets were deleted or removed, or a user changed their account access to private.

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

Dataset gathered and processed by Mark Dredze, Alexandra DeLucia, Shijie Wu, Aaron Mueller, Carlos Aguirre, and Philip Resnik.

Licensing Information

MIT

Citation Information

Please cite the Bernice paper if you use this dataset:

Alexandra DeLucia, Shijie Wu, Aaron Mueller, Carlos Aguirre, Philip Resnik, and Mark Dredze. 2022. Bernice: A Multilingual Pre-trained Encoder for Twitter. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 6191–6205, Abu Dhabi, United Arab Emirates. Association for Computational Linguistics.

Contributions

Dataset uploaded by @AADeLucia.

Downloads last month
181

Models trained or fine-tuned on jhu-clsp/bernice-pretrain-data