QARiB: QCRI Arabic and Dialectal BERT

About QARiB

QCRI Arabic and Dialectal BERT (QARiB) model, was trained on a collection of ~ 420 Million tweets and ~ 180 Million sentences of text. For tweets, the data was collected using twitter API and using language filter. lang:ar. For text data, it was a combination from Arabic GigaWord, Abulkhair Arabic Corpus and OPUS.

bert-base-qarib60_860k

• Data size: 60Gb
• Number of Iterations: 860k
• Loss: 2.2454472

Training QARiB

The training of the model has been performed using Google’s original Tensorflow code on Google Cloud TPU v2. We used a Google Cloud Storage bucket, for persistent storage of training data and models. See more details in Training QARiB

Using QARiB

You can use the raw model for either masked language modeling or next sentence prediction, but it's mostly intended to be fine-tuned on a downstream task. See the model hub to look for fine-tuned versions on a task that interests you. For more details, see Using QARiB

How to use

You can use this model directly with a pipeline for masked language modeling:

>>>from transformers import pipeline
>>>fill_mask = pipeline("fill-mask", model="./models/data60gb_86k")

>>> fill_mask("شو عندكم يا [MASK]")
[{'sequence': '[CLS] شو عندكم يا عرب [SEP]', 'score': 0.0990147516131401, 'token': 2355, 'token_str': 'عرب'}, 
{'sequence': '[CLS] شو عندكم يا جماعة [SEP]', 'score': 0.051633741706609726, 'token': 2308, 'token_str': 'جماعة'}, 
{'sequence': '[CLS] شو عندكم يا شباب [SEP]', 'score': 0.046871256083250046, 'token': 939, 'token_str': 'شباب'}, 
{'sequence': '[CLS] شو عندكم يا رفاق [SEP]', 'score': 0.03598872944712639, 'token': 7664, 'token_str': 'رفاق'}, 
{'sequence': '[CLS] شو عندكم يا ناس [SEP]', 'score': 0.031996358186006546, 'token': 271, 'token_str': 'ناس'}]

>>> fill_mask("قللي وشفيييك يرحم [MASK]")
[{'sequence': '[CLS] قللي وشفيييك يرحم والديك [SEP]', 'score': 0.4152909517288208, 'token': 9650, 'token_str': 'والديك'}, 
{'sequence': '[CLS] قللي وشفيييك يرحملي [SEP]', 'score': 0.07663793861865997, 'token': 294, 'token_str': '##لي'}, 
{'sequence': '[CLS] قللي وشفيييك يرحم حالك [SEP]', 'score': 0.0453166700899601, 'token': 2663, 'token_str': 'حالك'}, 
{'sequence': '[CLS] قللي وشفيييك يرحم امك [SEP]', 'score': 0.04390475153923035, 'token': 1942, 'token_str': 'امك'}, 
{'sequence': '[CLS] قللي وشفيييك يرحمونك [SEP]', 'score': 0.027349254116415977, 'token': 3283, 'token_str': '##ونك'}]

>>> fill_mask("وقام المدير [MASK]")
[
{'sequence': '[CLS] وقام المدير بالعمل [SEP]', 'score': 0.0678194984793663, 'token': 4230, 'token_str': 'بالعمل'}, 
{'sequence': '[CLS] وقام المدير بذلك [SEP]', 'score': 0.05191086605191231, 'token': 984, 'token_str': 'بذلك'}, 
{'sequence': '[CLS] وقام المدير بالاتصال [SEP]', 'score': 0.045264165848493576, 'token': 26096, 'token_str': 'بالاتصال'}, 
{'sequence': '[CLS] وقام المدير بعمله [SEP]', 'score': 0.03732728958129883, 'token': 40486, 'token_str': 'بعمله'}, 
{'sequence': '[CLS] وقام المدير بالامر [SEP]', 'score': 0.0246378555893898, 'token': 29124, 'token_str': 'بالامر'}
]
>>> fill_mask("وقامت المديرة [MASK]")

[{'sequence': '[CLS] وقامت المديرة بذلك [SEP]', 'score': 0.23992691934108734, 'token': 984, 'token_str': 'بذلك'}, 
{'sequence': '[CLS] وقامت المديرة بالامر [SEP]', 'score': 0.108805812895298, 'token': 29124, 'token_str': 'بالامر'}, 
{'sequence': '[CLS] وقامت المديرة بالعمل [SEP]', 'score': 0.06639821827411652, 'token': 4230, 'token_str': 'بالعمل'}, 
{'sequence': '[CLS] وقامت المديرة بالاتصال [SEP]', 'score': 0.05613093823194504, 'token': 26096, 'token_str': 'بالاتصال'}, 
{'sequence': '[CLS] وقامت المديرة المديرة [SEP]', 'score': 0.021778125315904617, 'token': 41635, 'token_str': 'المديرة'}]

Training procedure

The training of the model has been performed using Google’s original Tensorflow code on eight core Google Cloud TPU v2. We used a Google Cloud Storage bucket, for persistent storage of training data and models.

Eval results

We evaluated QARiB models on five NLP downstream task:

• Sentiment Analysis
• Emotion Detection
• Named-Entity Recognition (NER)
• Offensive Language Detection
• Dialect Identification

The results obtained from QARiB models outperforms multilingual BERT/AraBERT/ArabicBERT.

Model Weights and Vocab Download

From Huggingface site: https://huggingface.co/qarib/bert-base-qarib60_860k

Contacts

Ahmed Abdelali, Sabit Hassan, Hamdy Mubarak, Kareem Darwish and Younes Samih

Reference

@article{abdelali2021pretraining,
    title={Pre-Training BERT on Arabic Tweets: Practical Considerations},
    author={Ahmed Abdelali and Sabit Hassan and Hamdy Mubarak and Kareem Darwish and Younes Samih},
    year={2021},
    eprint={2102.10684},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}