CAMeLBERT is a collection of BERT models pre-trained on Arabic texts with different sizes and variants. We release pre-trained language models for Modern Standard Arabic (MSA), dialectal Arabic (DA), and classical Arabic (CA), in addition to a model pre-trained on a mix of the three. We also provide additional models that are pre-trained on a scaled-down set of the MSA variant (half, quarter, eighth, and sixteenth). The details are described in the paper "The Interplay of Variant, Size, and Task Type in Arabic Pre-trained Language Models."

This model card describes CAMeLBERT-Mix (bert-base-arabic-camelbert-mix), a model pre-trained on a mixture of these variants: MSA, DA, and CA.

	Model	Variant	Size	#Word
✔	`bert-base-arabic-camelbert-mix`	CA,DA,MSA	167GB	17.3B
	`bert-base-arabic-camelbert-ca`	CA	6GB	847M
	`bert-base-arabic-camelbert-da`	DA	54GB	5.8B
	`bert-base-arabic-camelbert-msa`	MSA	107GB	12.6B
	`bert-base-arabic-camelbert-msa-half`	MSA	53GB	6.3B
	`bert-base-arabic-camelbert-msa-quarter`	MSA	27GB	3.1B
	`bert-base-arabic-camelbert-msa-eighth`	MSA	14GB	1.6B
	`bert-base-arabic-camelbert-msa-sixteenth`	MSA	6GB	746M

Intended uses

You can use the released model for either masked language modeling or next sentence prediction. However, it is mostly intended to be fine-tuned on an NLP task, such as NER, POS tagging, sentiment analysis, dialect identification, and poetry classification. We release our fine-tuninig code here.

How to use

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

python

>>> from transformers import pipeline
>>> unmasker = pipeline('fill-mask', model='CAMeL-Lab/bert-base-arabic-camelbert-mix')
>>> unmasker("الهدف من الحياة هو [MASK] .")
[{'sequence': '[CLS] الهدف من الحياة هو النجاح. [SEP]',
  'score': 0.10861027985811234,
  'token': 6232,
  'token_str': 'النجاح'},
 {'sequence': '[CLS] الهدف من الحياة هو.. [SEP]',
  'score': 0.07626965641975403,
  'token': 18,
  'token_str': '.'},
 {'sequence': '[CLS] الهدف من الحياة هو الحياة. [SEP]',
  'score': 0.05131986364722252,
  'token': 3696,
  'token_str': 'الحياة'},
 {'sequence': '[CLS] الهدف من الحياة هو الموت. [SEP]',
  'score': 0.03734956309199333,
  'token': 4295,
  'token_str': 'الموت'},
 {'sequence': '[CLS] الهدف من الحياة هو العمل. [SEP]',
  'score': 0.027189988642930984,
  'token': 2854,
  'token_str': 'العمل'}]

Note: to download our models, you would need transformers>=3.5.0. Otherwise, you could download the models manually.

Here is how to use this model to get the features of a given text in PyTorch:

python

from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained('CAMeL-Lab/bert-base-arabic-camelbert-mix')
model = AutoModel.from_pretrained('CAMeL-Lab/bert-base-arabic-camelbert-mix')
text = "مرحبا يا عالم."
encoded_input = tokenizer(text, return_tensors='pt')
output = model(**encoded_input)

and in TensorFlow:

python

from transformers import AutoTokenizer, TFAutoModel
tokenizer = AutoTokenizer.from_pretrained('CAMeL-Lab/bert-base-arabic-camelbert-mix')
model = TFAutoModel.from_pretrained('CAMeL-Lab/bert-base-arabic-camelbert-mix')
text = "مرحبا يا عالم."
encoded_input = tokenizer(text, return_tensors='tf')
output = model(encoded_input)

Training data

MSA (Modern Standard Arabic)
- The Arabic Gigaword Fifth Edition
- Abu El-Khair Corpus
- OSIAN corpus
- Arabic Wikipedia
- The unshuffled version of the Arabic OSCAR corpus
DA (dialectal Arabic)
- A collection of dialectal Arabic data described in our paper.
CA (classical Arabic)
- OpenITI (Version 2020.1.2)

Training procedure

We use the original implementation released by Google for pre-training. We follow the original English BERT model's hyperparameters for pre-training, unless otherwise specified.

Preprocessing

After extracting the raw text from each corpus, we apply the following pre-processing.
We first remove invalid characters and normalize white spaces using the utilities provided by the original BERT implementation.
We also remove lines without any Arabic characters.
We then remove diacritics and kashida using CAMeL Tools.
Finally, we split each line into sentences with a heuristics-based sentence segmenter.
We train a WordPiece tokenizer on the entire dataset (167 GB text) with a vocabulary size of 30,000 using HuggingFace's tokenizers.
We do not lowercase letters nor strip accents.

Pre-training

The model was trained on a single cloud TPU (v3-8) for one million steps in total.
The first 90,000 steps were trained with a batch size of 1,024 and the rest was trained with a batch size of 256.
The sequence length was limited to 128 tokens for 90% of the steps and 512 for the remaining 10%.
We use whole word masking and a duplicate factor of 10.
We set max predictions per sequence to 20 for the dataset with max sequence length of 128 tokens and 80 for the dataset with max sequence length of 512 tokens.
We use a random seed of 12345, masked language model probability of 0.15, and short sequence probability of 0.1.
The optimizer used is Adam with a learning rate of 1e-4, \(\beta_{1} = 0.9\) and \(\beta_{2} = 0.999\), a weight decay of 0.01, learning rate warmup for 10,000 steps and linear decay of the learning rate after.

Evaluation results

We evaluate our pre-trained language models on five NLP tasks: NER, POS tagging, sentiment analysis, dialect identification, and poetry classification.
We fine-tune and evaluate the models using 12 dataset.
We used Hugging Face's transformers to fine-tune our CAMeLBERT models.
We used transformers v3.1.0 along with PyTorch v1.5.1.
The fine-tuning was done by adding a fully connected linear layer to the last hidden state.
We use \(F_{1}\) score as a metric for all tasks.
Code used for fine-tuning is available here.

Results

Task	Dataset	Variant	Mix	CA	DA	MSA	MSA-1/2	MSA-1/4	MSA-1/8	MSA-1/16
NER	ANERcorp	MSA	80.8%	67.9%	74.1%	82.4%	82.0%	82.1%	82.6%	80.8%
POS	PATB (MSA)	MSA	98.1%	97.8%	97.7%	98.3%	98.2%	98.3%	98.2%	98.2%
	ARZTB (EGY)	DA	93.6%	92.3%	92.7%	93.6%	93.6%	93.7%	93.6%	93.6%
	Gumar (GLF)	DA	97.3%	97.7%	97.9%	97.9%	97.9%	97.9%	97.9%	97.9%
SA	ASTD	MSA	76.3%	69.4%	74.6%	76.9%	76.0%	76.8%	76.7%	75.3%
	ArSAS	MSA	92.7%	89.4%	91.8%	93.0%	92.6%	92.5%	92.5%	92.3%
	SemEval	MSA	69.0%	58.5%	68.4%	72.1%	70.7%	72.8%	71.6%	71.2%
DID	MADAR-26	DA	62.9%	61.9%	61.8%	62.6%	62.0%	62.8%	62.0%	62.2%
	MADAR-6	DA	92.5%	91.5%	92.2%	91.9%	91.8%	92.2%	92.1%	92.0%
	MADAR-Twitter-5	MSA	75.7%	71.4%	74.2%	77.6%	78.5%	77.3%	77.7%	76.2%
	NADI	DA	24.7%	17.3%	20.1%	24.9%	24.6%	24.6%	24.9%	23.8%
Poetry	APCD	CA	79.8%	80.9%	79.6%	79.7%	79.9%	80.0%	79.7%	79.8%

Results (Average)

	Variant	Mix	CA	DA	MSA	MSA-1/2	MSA-1/4	MSA-1/8	MSA-1/16
Variant-wise-average^[1]	MSA	82.1%	75.7%	80.1%	83.4%	83.0%	83.3%	83.2%	82.3%
	DA	74.4%	72.1%	72.9%	74.2%	74.0%	74.3%	74.1%	73.9%
	CA	79.8%	80.9%	79.6%	79.7%	79.9%	80.0%	79.7%	79.8%
Macro-Average	ALL	78.7%	74.7%	77.1%	79.2%	79.0%	79.2%	79.1%	78.6%

[1]: Variant-wise-average refers to average over a group of tasks in the same language variant.

Acknowledgements

This research was supported with Cloud TPUs from Google’s TensorFlow Research Cloud (TFRC).

Citation

bibtex

@inproceedings{inoue-etal-2021-interplay,
    title = "The Interplay of Variant, Size, and Task Type in {A}rabic Pre-trained Language Models",
    author = "Inoue, Go  and
      Alhafni, Bashar  and
      Baimukan, Nurpeiis  and
      Bouamor, Houda  and
      Habash, Nizar",
    booktitle = "Proceedings of the Sixth Arabic Natural Language Processing Workshop",
    month = apr,
    year = "2021",
    address = "Kyiv, Ukraine (Online)",
    publisher = "Association for Computational Linguistics",
    abstract = "In this paper, we explore the effects of language variants, data sizes, and fine-tuning task types in Arabic pre-trained language models. To do so, we build three pre-trained language models across three variants of Arabic: Modern Standard Arabic (MSA), dialectal Arabic, and classical Arabic, in addition to a fourth language model which is pre-trained on a mix of the three. We also examine the importance of pre-training data size by building additional models that are pre-trained on a scaled-down set of the MSA variant. We compare our different models to each other, as well as to eight publicly available models by fine-tuning them on five NLP tasks spanning 12 datasets. Our results suggest that the variant proximity of pre-training data to fine-tuning data is more important than the pre-training data size. We exploit this insight in defining an optimized system selection model for the studied tasks.",
}