aubmindlab/bert-large-arabertv2

AraBERT v1 & v2 : Pre-training BERT for Arabic Language Understanding

AraBERT is an Arabic pretrained language model based on Google's BERT architechture. AraBERT uses the same BERT-Base config. More details are available in the AraBERT Paper and in the AraBERT Meetup

There are two versions of the model, AraBERTv0.1 and AraBERTv1, with the difference being that AraBERTv1 uses pre-segmented text where prefixes and suffixes were split using the Farasa Segmenter.

We evaluate AraBERT models on different downstream tasks and compare them to mBERT, and other state of the art models (To the extent of our knowledge). The Tasks were Sentiment Analysis on 6 different datasets (HARD, ASTD-Balanced, ArsenTD-Lev, LABR), Named Entity Recognition with the ANERcorp, and Arabic Question Answering on Arabic-SQuAD and ARCD

AraBERTv2

What's New!

AraBERT now comes in 4 new variants to replace the old v1 versions:

More Detail in the AraBERT folder and in the README and in the AraBERT Paper

Model	HuggingFace Model Name	Size (MB/Params)	Pre-Segmentation	DataSet (Sentences/Size/nWords)
AraBERTv0.2-base	bert-base-arabertv02	543MB / 136M	No	200M / 77GB / 8.6B
AraBERTv0.2-large	bert-large-arabertv02	1.38G 371M	No	200M / 77GB / 8.6B
AraBERTv2-base	bert-base-arabertv2	543MB 136M	Yes	200M / 77GB / 8.6B
AraBERTv2-large	bert-large-arabertv2	1.38G 371M	Yes	200M / 77GB / 8.6B
AraBERTv0.2-Twitter-base	bert-base-arabertv02-twitter	543MB / 136M	No	Same as v02 + 60M Multi-Dialect Tweets
AraBERTv0.2-Twitter-large	bert-large-arabertv02-twitter	1.38G / 371M	No	Same as v02 + 60M Multi-Dialect Tweets
AraBERTv0.1-base	bert-base-arabertv01	543MB 136M	No	77M / 23GB / 2.7B
AraBERTv1-base	bert-base-arabert	543MB 136M	Yes	77M / 23GB / 2.7B

All models are available in the HuggingFace model page under the aubmindlab name. Checkpoints are available in PyTorch, TF2 and TF1 formats.

Better Pre-Processing and New Vocab

We identified an issue with AraBERTv1's wordpiece vocabulary. The issue came from punctuations and numbers that were still attached to words when learned the wordpiece vocab. We now insert a space between numbers and characters and around punctuation characters.

The new vocabulary was learned using the BertWordpieceTokenizer from the tokenizers library, and should now support the Fast tokenizer implementation from the transformers library.

P.S.: All the old BERT codes should work with the new BERT, just change the model name and check the new preprocessing function Please read the section on how to use the preprocessing function

Bigger Dataset and More Compute

We used ~3.5 times more data, and trained for longer. For Dataset Sources see the Dataset Section

Model	Hardware	num of examples with seq len (128 / 512)	128 (Batch Size/ Num of Steps)	512 (Batch Size/ Num of Steps)	Total Steps	Total Time (in Days)
AraBERTv0.2-base	TPUv3-8	420M / 207M	2560 / 1M	384/ 2M	3M	-
AraBERTv0.2-large	TPUv3-128	420M / 207M	13440 / 250K	2056 / 300K	550K	7
AraBERTv2-base	TPUv3-8	420M / 207M	2560 / 1M	384/ 2M	3M	-
AraBERTv2-large	TPUv3-128	520M / 245M	13440 / 250K	2056 / 300K	550K	7
AraBERT-base (v1/v0.1)	TPUv2-8	-	512 / 900K	128 / 300K	1.2M	4

Dataset

The pretraining data used for the new AraBERT model is also used for Arabic GPT2 and ELECTRA.

The dataset consists of 77GB or 200,095,961 lines or 8,655,948,860 words or 82,232,988,358 chars (before applying Farasa Segmentation)

For the new dataset we added the unshuffled OSCAR corpus, after we thoroughly filter it, to the previous dataset used in AraBERTv1 but with out the websites that we previously crawled:

• OSCAR unshuffled and filtered.
• Arabic Wikipedia dump from 2020/09/01
• The 1.5B words Arabic Corpus
• The OSIAN Corpus
• Assafir news articles. Huge thank you for Assafir for providing us the data

Preprocessing

It is recommended to apply our preprocessing function before training/testing on any dataset.

Install the arabert python package to segment text for AraBERT v1 & v2 or to clean your data pip install arabert

from arabert.preprocess import ArabertPreprocessor

model_name="aubmindlab/bert-large-arabertv2"
arabert_prep = ArabertPreprocessor(model_name=model_name)

text = "ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري"
arabert_prep.preprocess(text)
>>>"و+ لن نبالغ إذا قل +نا إن هاتف أو كمبيوتر ال+ مكتب في زمن +نا هذا ضروري"

TensorFlow 1.x models

The TF1.x model are available in the HuggingFace models repo. You can download them as follows:

• via git-lfs: clone all the models in a repo

curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
sudo apt-get install git-lfs
git lfs install
git clone https://huggingface.co/aubmindlab/MODEL_NAME
tar -C ./MODEL_NAME -zxvf /content/MODEL_NAME/tf1_model.tar.gz

where MODEL_NAME is any model under the aubmindlab name

• via wget:
  • Go to the tf1_model.tar.gz file on huggingface.co/models/aubmindlab/MODEL_NAME.
  • copy the oid sha256
  • then run wget https://cdn-lfs.huggingface.co/aubmindlab/aragpt2-base/INSERT_THE_SHA_HERE (ex: for aragpt2-base: wget https://cdn-lfs.huggingface.co/aubmindlab/aragpt2-base/3766fc03d7c2593ff2fb991d275e96b81b0ecb2098b71ff315611d052ce65248)

If you used this model please cite us as :

Google Scholar has our Bibtex wrong (missing name), use this instead

@inproceedings{antoun2020arabert,
  title={AraBERT: Transformer-based Model for Arabic Language Understanding},
  author={Antoun, Wissam and Baly, Fady and Hajj, Hazem},
  booktitle={LREC 2020 Workshop Language Resources and Evaluation Conference 11--16 May 2020},
  pages={9}
}

Acknowledgments

Thanks to TensorFlow Research Cloud (TFRC) for the free access to Cloud TPUs, couldn't have done it without this program, and to the AUB MIND Lab Members for the continuous support. Also thanks to Yakshof and Assafir for data and storage access. Another thanks for Habib Rahal (https://www.behance.net/rahalhabib), for putting a face to AraBERT.

Contacts

Wissam Antoun: Linkedin | Twitter | Github | [email protected] | [email protected]

Fady Baly: Linkedin | Twitter | Github | [email protected] | [email protected]