ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic

What is the repository is about?

This is the repository accompanying our ACL 2021 paper ARBERT & MARBERT: Deep Bidirectional Transformers for Arabic. In the paper, we:

• introduce ARBERT and MARBERT, two powerful Transformer-based language models for Arabic;
• introduce ArBench, a multi-domain, multi-variety benchmark for Arabic naturaal language understanding based on 41 datasets across 5 different tasks and task clusters;
• evaluate ARBERT and MARBERT on ArBench and compare against available language models.

Our models establish new state-of-the-art (SOTA) on all 5 tasks/task clusters on 37 out of the 41 datasets. Our language models are publicaly available for research (see below). The rest of this repository provides more information about our new language models, benchmark, and experiments.

---

Table of Contents

• 1 Our Language Models
  • 1.1 ARBERT & MARBERT
  • 1.2 Training Data and Vocabulary
• 2. Our Benchmark: ArBench
  • 2.1 Sentiment Analysis
  • 2.2 Social Meaning
  • 2.3 Topic Classification
  • 2.4 Dialect Identification
  • 2.5 Named Entity Recogntion (NER)
• 3. Model Evaluation
  • 3.1 Sentiment Analysis
  • 3.2 Social Meaning
  • 3.3 Topic Classification
  • 3.4 Dialect Identification
  • 3.5 Named Entity Recogntion (NER)
• 4. How to use ARBERT and MARBERT
• 5. Ethics
• 6. Download ARBERT and MARBERT Checkpoints
• 7. Citation
• 8. Acknowledgments

---

1. Our Language Models

1.1 ARBERT & MARBERT

ARBERT is a large scale pre-training masked language model focused on Modern Standard Arabic (MSA). To train ARBERT, we use the same architecture as BERT-base: 12 attention layers, each has 12 attention heads and 768 hidden dimensions, a vocabulary of 100K WordPieces, making ∼163M parameters. We train ARBERT on a collection of Arabic datasets comprising 61GB of text (6.2B tokens)

MARBERT is a large scale pre-training masked language model focused on both Dialectal Arabic (DA) and MSA. Arabic has multiple varieties. To train MARBERT, we randomly sample 1B Arabic tweets from a large in-house dataset of about 6B tweets. We only include tweets with at least 3 Arabic words, based on character string matching, regardless whether the tweet has non-Arabic string or not. That is, we do not remove non-Arabic so long as the tweet meets the 3 Arabic word criterion. The dataset makes up 128GB of text (15.6B tokens). We use the same network architecture as ARBERT (BERT-base), but without the next sentence prediction (NSP) objective since tweets are short. See our repo for modifying BERT code to remove NSP.

1.2 Training Data and Vocabulary

The following table shows a comparison between ARBERT and MARBERT, on the one hand, and mBERT, XLM-R, and AraBERT, on the other hand. We compare in terms of pre-training data sources and size, vocabulary size, and model parameter size.

	Data Source	#Tokens(ar/all)	Tokanization	Vocab Size(ar/all)	Cased	Arch.	#Param
mBERT	Wikipedia	153M/1.5B	WordPiece	5K/110K	yes	base	110M
XLM-RB	CommonCraw	l2.9B/295B	SentencePiece	14K/250K	yes	base	270M
XLM-RL	CommonCraw	l2.9B/295B	SentencePiece	14K/250K	yes	large	550M
AraBERT	Several (3 sources)	2.5B/2.5B	SentencePiece	60K/64K	no	base	135M
ARBERT	Several (6 sources)	6.2B/6.2B	WordPiece	100K/100K	no	base	163M
MARBERT	Arabic Twitter	15.6B/15.6B	WordPiece	100K/100K	no	base	163M

---

2. Our Benchmark: ArBench

To evaluate our models, we also introduce ArBench, a new benchmark for multi-dialectal Arabic language understanding. ArBench is built using 41 datasets targeting 5 different tasks/task clusters, allowing us to offer a series of standardized experiments under rich conditions. The following are the different tasks/task clusers covered by ArBench:

2.1 Sentiment Analysis

Reference	Data (#classes)	TRAIN	DEV	TEST
Alomari et al. (2017)	AJGT (2)	1.4K	-	361
Abdul-Mageed et al. (2020b)	AraNETSent (2)	100K	14.3K	11.8K
Al-Twairesh et al. (2017)	AraSenTi (3)	11,117	1,407	1,382
Abu Farha and Magdy (2017)	ArSarcasmSent (3)	8.4K	-	2.K
Elmadany et al. (2018)	ArSAS (3)	24.7K	-	3.6K
Baly et al. (2019)	ArsenTD-LEV (5)	3.2K	-	801
Nabil et al. (2015)	ASTD (3)	24.7K	-	664
Nabil et al. (2015)	ASTD-B (2)	1.06K	-	267
Abdul-Mageed and Diab (2012)	AWATIF (4)	2.28K	288	284
Salameh et al. (2015)	BBN (3)	960	125	116
Elnagar et al. (2018)	HARD (2)	84.5K	-	21.1K
Nabil et al. (2015)	LABR (2)	13.1K		3.28K
Abdul-Mageed and Diab (2014)	SAMAR (5)	2.49K	310	316
Rosenthal et al. (2017)	SemEval (3)	24.7K	-	6.10K
Salameh et al. (2015)	SYTS(3)	960	202	199
Saad (2019)	TwitterSaad (2)	1.5K	202	190
Abdullah et al. (2013)	TwitterAbdullah (2)	46k	5.77k	5.82k

2.2 Social Meaning

Reference	Task	Data (#classes)	TRAIN	DEV	TEST
Zaghouani and Charfi (2018)	Age	Arap-Tweet (3)	1.28M	160K	160K
Zaghouani and Charfi (2018)	Gender	Arap-Tweet (2)	1.28M	160K	160K
Abdul-Mageed et al. (2020b)	Emotion	AraNETEmo (8)	189K	911	942
Abu Farha and Magdy (2017)	Sarcasm	AraSarcasm (2)	8.4K	-	2.1K
Alshehri et al. (2020a)	Dangerous	AraDang (2)	3.4K	616	664
Ghanem et al. (2019)	Irony	FIRE2019 (2)	3.6K	-	404
Mubarak et al. (2020)	Offensive	OSACT-A (2)	10K	1K	2K
Mubarak et al. (2020)	Hate Speech	OSACT-B - (2)	10K	1K	2K

2.3 Topic Classification

Reference	Data (#classes)	TRAIN	DEV	TEST
Saad and Ashour (2010)	OSAC (10)	17.9K	2.24K	2.24K
Abbas et al. (2011)	Khallej (4)	4.55K	570	570
Chouigui et al. (2017)	ANT(5)	25.2K	31.5K	31.5K

2.4 Dialect Identification

Reference	Data (#classes)	Task	TRAIN	DEV	TEST
Zaidan and Callison-Burch (2014)	AOC (2)	Binary	86.5K	10.8K	10.8K
Zaidan and Callison-Burch (2014)	AOC (3)	Region	35.7K	4.46K	4.45K
Zaidan and Callison-Burch (2014)	AOC (4)	Region	86.5K	10.8K	10.8K
Farha and Magdy (2020)	ArSarcasmDia (5)	Regoin	8.43K	-	2.11K
Bouamor et al. (2019)	MADAR-TL (21)	Country	193K	26.6K	43.9K
Abdul-Mageed et al. (2020)	NADI (21)	Country	2.1K	4.96K	5K
Abdul-Mageed et al. (2020)	NADI (100)	Province	2.1K	4.96K	5K
Abdelali et al. (2020)	QADI (18)	Country	498K	-	3.5K

2.5 Named Entity Recogntion

Reference	Dataset	#Tokens	#PER	#LOC	#ORG
Benajiba and Rosso (2007)	ANERCorp	150K	6.50K	5.01K	3.43K
Mitchell et al. (2004)	ACE-2003BN	15K	832	1.22K	288
Mitchell et al. (2004)	ACE-2003NW	27K	1.14K	1.14K	893
Mitchell et al. (2005)	ACE-2004BN	70K	3.20K	3.92K	2.23K
Darwish (2013)	TW-NER	81K	1.25K	1.30K	765

---

3 Model Evaluation

When fine-tuned on ArBench, ARBERT and MARBERT collectively achieve new SOTA with sizeable margins compared to all existing models such as mBERT, XLM-R (Base and Large), and AraBERT on 37 out of 45 classification tasks on the 41 datasets (82.22%). We present our results on the different TEST sets in the subsections below. For performance on DEV sets, please see appendixes in our paper.

3.1 Sentiment Analysis

Dataset (#classes)	mBERT	XLM-RB	XLM-RL	AraBERT	ARBERT	MARBERT
AJGT (2)	86.67	89.44	91.94	92.22	94.44	96.11
HARD (2)	95.54	95.74	95.96	95.89	96.12	96.17
ArsenTD-LEV (5)	50.50	55.25	62.00	56.13	61.38	60.38
LABR (2)	91.20	91.23	92.20	91.97	92.51	92.49
ASTD-B(2)	79.32	87.59	77.44	83.08	93.23	96.24

Results reported based on Acc. score

Dataset (#classes)	mBERT	XLM-RB	XLM-RL	AraBERT	ARBERT	MARBERT
ArSAS (3)	87.50	90.00	91.50	91.00	92.00	93.00
ASTD (3)	67.00	60.67	67.67	72.00	76.50	78.00
SemEval (3)	57.00	64.00	67.00	62.00	69.00	71.00
AraNETSent (2)	84.00	92.00	93.00	86.50	89.00	92.00
ArSarcasmSent (3)	60.50	63.50	70.00	63.50	68.00	71.50
AraSenTi (noura) (3)	89.50	92.00	93.50	91.00	90.00	90.00
BBN(3)	55.50	69.50	46.50	70.00	76.50	79.00
SYTS(3)	67.00	78.00	40.50	75.50	79.00	76.50
TwitterSaad (2)	79.00	95.00	95.00	81.00	90.00	96.00
SAMAR(5)	22.50	54.00	57.00	36.50	43.50	55.50
AWATIF(4)	60.50	63.50	68.50	66.50	71.50	72.50
TwitterAbdullah (2)	81.50	91.00	92.00	89.50	91.50	95.00

Results reported based on F₁^NP score.

3.2 Social Meaning

Task (#classes)	Dataset	mBERT	XLM-RB	XLM-RL	AraBERT	ARBERT	MARBERT
Offensive (2)	OSACT-A	84.25	85.26	88.28	86.57	90.38	92.41
Hate Speech(2)	OSACT-B	72.81	71.33	79.31	78.89	83.02	84.79
Dangerous (2)	Dangerous	62.66	62.76	65.01	64.37	63.21	67.53
Sarcasm (2)	AraSarcasm	68.20	66.76	69.23	72.23	75.04	76.30
Emotion (8)	AraNETEmo	65.79	70.67	74.89	65.68	67.73	75.83
Irony (2)	FIRE2019	80.96	81.97	82.52%	83.01	85.59	85.33
Age (3)	Arab_Tweet	56.35	59.73	53.60	57.72	58.95	62.27
Gender (2)	Arab_Tweet	68.06	71.00	71.14	67.75	69.86	72.62

Results reported based on F₁ score.

3.3 Topic Classification

Dataset (#classes)	mBERT	XLM-RB	XLM-RL	AraBERT	ARBERT	MARBERT
OSAC (10)	96.84	97.15	98.20	97.03	97.50	97.23
Khallej (4)	92.81	91.87	93.56	93.83	94.53	93.63
ANTText (5)	84.89	85.77	86.72	88.17	86.87	85.27
ANTTitle (5)	78.29	79.96	81.25	81.03	81.70	81.19
ANTText+Title (5)	84.67	86.21	86.96	87.22	87.21	85.60

Results reported based on F₁ score.

3.4 Dialect Identification

Task (#classes)	Dataset	mBERT	XLM-RB	XLM-RL	AraBERT	ARBERT	MARBERT
Regoin (5)	ArSarcasmDia	43.81	41.71	41.83	47.54	51.27	54.70
Country (21)	MADAR-TL	34.92	35.91	35.14	34.87	37.90	40.40
Region (4)	AOC	77.27	77.34	78.77	79.20	81.09	82.37
Region (3)	AOC	85.76	86.39	87.56	87.68	89.06	90.85
Binary (4)	AOC	86.19	86.85	87.30	87.76	88.46	88.59
Country(18)	QADI	66.57	77.00	82.73	72.23	88.63	90.89
Country(21)	NADI	13.32	16.36	17.17	17.46	22.56	29.14
Province (100 )	NADI	2.13	4.12	0.32	3.13	6.10	6.28

Results reported based on F₁ score.

3.5 Named Entity Recogntion

Dataset	mBERT	XLM-RB	XLM-RL	AraBERT	ARBERT	MARBERT
ANERcorp.	86.78	87.24	89.94	89.13	84.38	80.64
ACE 2004 NW	86.37	89.93	89.89	89.03	88.24	85.02
ACE 2003BN	91.23	53.97	85.41	91.94	96.18	79.05
ACE 2003NW	81.40	87.24	90.62	88.09	90.09	87.76
TW-NER	36.83	49.16	54.44	41.26	59.17	67.39

Results reported based on F₁ score.

---

4. How to use ARBERT and MARBERT

4.1 Loading directly from Huggingface

You can use ARBERT and MARBERT with Hugging Face's Transformers library as follow.

   from transformers import AutoTokenizer, AutoModel
   #load AEBERT model from huggingface
   ARBERT_tokenizer = AutoTokenizer.from_pretrained("UBC-NLP/ARBERT")
   ARBERT_model = AutoModel.from_pretrained("UBC-NLP/ARBERT")
 
   #load MAEBERT model from huggingface
   MARBERT_tokenizer = AutoTokenizer.from_pretrained("UBC-NLP/MARBERT")
   MARBERT_model = AutoModel.from_pretrained("UBC-NLP/MARBERT") 

4.2 Example: Fine-tuning MARBERT for Sentiment Analysis

MARBERT Fine-Tuning demo with PyTorch checkpoint for Sentiment Analysis on the AJGT dataset

5. Ethics

Our models are developed using data from the public domain. We provide access to our models to accelerate scientific research with no liability on our part. Please use our models and benchmark only ethically. This includes, for example, respect and protection of people's privacy. We encourage all researchers who decide to use our models to adhere to the highest standards. For example, if you apply our models on Twitter data, we encourage you to review Twitter policy at Twitter policy. For example, Twitter provides the following policy around use of sensitive information:

Sensitive information

You should be careful about using Twitter data to derive or infer potentially sensitive characteristics about Twitter users. Never derive or infer, or store derived or inferred, information about a Twitter user’s:

• Health (including pregnancy)
• Negative financial status or condition
• Political affiliation or beliefs
• Racial or ethnic origin
• Religious or philosophical affiliation or beliefs
• Sex life or sexual orientation
• Trade union membership
• Alleged or actual commission of a crime
• Aggregate analysis of Twitter content that does not store any personal data (for example, user IDs, usernames, and other identifiers) is permitted, provided that the analysis also complies with applicable laws and all parts of the Developer Agreement and Policy.

---

6. Download ARBERT and MARBERT Checkpoints

ARBERT and MARBERT Pytorch and Tenserflow checkpoints are available on Huggingface website for direct download and use exclusively for research. For commercial use, please contact the authors via email @ (*muhammad.mageed[at]ubc[dot]ca*).

Model	Pytorch Checkpoint	Tensorflow Checkpoint
ARBERT	Download	Download
MARBERT	Download	Download

---

7. Citation

If you use our models (ARBERT or MARBERT) for your scientific publication, or if you find the resources in this repository useful, please cite our paper as follows (to be updated):

@inproceedings{abdul-mageed-etal-2021-arbert,
    title = "{ARBERT} {\&} {MARBERT}: Deep Bidirectional Transformers for {A}rabic",
    author = "Abdul-Mageed, Muhammad  and
      Elmadany, AbdelRahim  and
      Nagoudi, El Moatez Billah",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.551",
    doi = "10.18653/v1/2021.acl-long.551",
    pages = "7088--7105",
    abstract = "Pre-trained language models (LMs) are currently integral to many natural language processing systems. Although multilingual LMs were also introduced to serve many languages, these have limitations such as being costly at inference time and the size and diversity of non-English data involved in their pre-training. We remedy these issues for a collection of diverse Arabic varieties by introducing two powerful deep bidirectional transformer-based models, ARBERT and MARBERT. To evaluate our models, we also introduce ARLUE, a new benchmark for multi-dialectal Arabic language understanding evaluation. ARLUE is built using 42 datasets targeting six different task clusters, allowing us to offer a series of standardized experiments under rich conditions. When fine-tuned on ARLUE, our models collectively achieve new state-of-the-art results across the majority of tasks (37 out of 48 classification tasks, on the 42 datasets). Our best model acquires the highest ARLUE score (77.40) across all six task clusters, outperforming all other models including XLM-R Large ( 3.4x larger size). Our models are publicly available at https://github.com/UBC-NLP/marbert and ARLUE will be released through the same repository.",
}

---

8. Acknowledgments

We gratefully acknowledge support from the Natural Sciences and Engineering Research Council of Canada, the Social Sciences and Humanities Research Council of Canada, Canadian Foundation for Innovation, ComputeCanada and UBC ARC-Sockeye. We also thank the Google TensorFlow Research Cloud (TFRC) program for providing us with free TPU access.