mbert_c19: An mbert model pretrained on 1.5 million COVID-19 multi-dialect Arabic tweets

mBERT COVID-19 Arxiv URL is a pretrained (fine-tuned) version of the mBERT model (https://huggingface.co/bert-base-multilingual-cased). The pretraining was done using 1.5 million multi-dialect Arabic tweets regarding the COVID-19 pandemic from the “Large Arabic Twitter Dataset on COVID-19” (https://arxiv.org/abs/2004.04315). The model can achieve better results for the tasks that deal with multi-dialect Arabic tweets in relation to the COVID-19 pandemic.

Classification results for multiple tasks including fake-news and hate speech detection when using arabert_c19 and mbert_ar_c19:

For more details refer to the paper (link)

	arabert	mbert	distilbert multi	arabert Covid-19	mbert Covid-19
Contains hate (Binary)	0.8346	0.6675	0.7145	`0.8649`	0.8492
Talk about a cure (Binary)	0.8193	0.7406	0.7127	0.9055	`0.9176`
News or opinion (Binary)	0.8987	0.8332	0.8099	`0.9163`	0.9116
Contains fake information (Binary)	0.6415	0.5428	0.4743	`0.7739`	0.7228

Preprocessing

from arabert.preprocess import ArabertPreprocessor
model_name="moha/mbert_ar_c19"
arabert_prep = ArabertPreprocessor(model_name=model_name)
text = "للوقايه من عدم انتشار كورونا عليك اولا غسل اليدين بالماء والصابون وتكون عملية الغسل دقيقه تشمل راحة اليد الأصابع التركيز على الإبهام"
arabert_prep.preprocess(text)

Citation

Please cite as:

@misc{ameur2021aracovid19mfh,
      title={AraCOVID19-MFH: Arabic COVID-19 Multi-label Fake News and Hate Speech Detection Dataset}, 
      author={Mohamed Seghir Hadj Ameur and Hassina Aliane},
      year={2021},
      eprint={2105.03143},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contacts

Hadj Ameur: Github | mohamedhadjameur@gmail.com | mhadjameur@cerist.dz