File size: 5,584 Bytes
e4ea8a9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f6259b1
 
e4ea8a9
f6259b1
 
 
 
 
 
e4ea8a9
 
 
f6259b1
 
 
 
 
e4ea8a9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
---

language: ar
datasets:
 - wikipedia
 - OSIAN
 - 1.5B Arabic Corpus
 - OSCAR Arabic Unshuffled
---


# AraELECTRA

**ELECTRA** is a method for self-supervised language representation learning. It can be used to pre-train transformer networks using relatively little compute. ELECTRA models are trained to distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to the discriminator of a [GAN](https://arxiv.org/pdf/1406.2661.pdf).  AraELECTRA achieves state-of-the-art results on Arabic QA dataset.

For a detailed description, please refer to the AraELECTRA paper [AraELECTRA: Pre-Training Text Discriminators for Arabic Language Understanding](https://arxiv.org/abs/2012.15516).

## How to use the discriminator in `transformers`

```python

from transformers import ElectraForPreTraining, ElectraTokenizerFast

import torch



discriminator = ElectraForPreTraining.from_pretrained("aubmindlab/araelectra-base-discriminator")

tokenizer = ElectraTokenizerFast.from_pretrained("aubmindlab/araelectra-base-discriminator")



sentence = ""

fake_sentence = ""



fake_tokens = tokenizer.tokenize(fake_sentence)

fake_inputs = tokenizer.encode(fake_sentence, return_tensors="pt")

discriminator_outputs = discriminator(fake_inputs)

predictions = torch.round((torch.sign(discriminator_outputs[0]) + 1) / 2)



[print("%7s" % token, end="") for token in fake_tokens]



[print("%7s" % int(prediction), end="") for prediction in predictions.tolist()]

```

# Model

Model | HuggingFace Model Name | Size (MB/Params)|
 ---|:---:|:---:
AraELECTRA-base-generator | [araelectra-base-generator](https://huggingface.co/aubmindlab/araelectra-base-generator) |  227MB/60M   |
AraELECTRA-base-discriminator | [araelectra-base-discriminator](https://huggingface.co/aubmindlab/araelectra-base-discriminator) |  516MB/135M  |

# Compute
Model | Hardware | num of examples (seq len = 512) | Batch Size | Num of Steps | Time (in days)
 ---|:---:|:---:|:---:|:---:|:---:
AraELECTRA-base | TPUv3-8 | - | 256 | 2M | 24

# Dataset

The pretraining data used for the new **AraELECTRA** model is also used for **AraGPT2 and AraBERTv2**.

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](https://archive.org/details/arwiki-20190201) from 2020/09/01
- [The 1.5B words Arabic Corpus](https://www.semanticscholar.org/paper/1.5-billion-words-Arabic-Corpus-El-Khair/f3eeef4afb81223df96575adadf808fe7fe440b4)
- [The OSIAN Corpus](https://www.aclweb.org/anthology/W19-4619)
- Assafir news articles. Huge thank you for Assafir for giving us the data

# Preprocessing

It is recommended to apply our preprocessing function before training/testing on any dataset.
**Install farasapy to segment text for AraBERT v1 & v2 `pip install farasapy`**

```python

from arabert.preprocess import ArabertPreprocessor



model_name="araelectra-base"

arabert_prep = ArabertPreprocessor(model_name=model_name, keep_emojis=True)



text = "ولن نبالغ إذا قلنا إن هاتف أو كمبيوتر المكتب في زمننا هذا ضروري"

arabert_prep.preprocess(text)

```


# TensorFlow 1.x models

The TF1.x model are avaiable in the HuggingFace models repo.
You can download them as follows:
- via git-lfs: clone all the models in a repo
```bash

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 :

```

@misc{antoun2020aragpt2,

      title={AraGPT2: Pre-Trained Transformer for Arabic Language Generation},

      author={Wissam Antoun and Fady Baly and Hazem Hajj},

      year={2020},

      eprint={2012.15520},

      archivePrefix={arXiv},

      primaryClass={cs.CL}

}

```

# 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](https://sites.aub.edu.lb/mindlab/) Members for the continous support. Also thanks to [Yakshof](https://www.yakshof.com/#/) 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](https://www.linkedin.com/in/wissam-antoun-622142b4/) | [Twitter](https://twitter.com/wissam_antoun) | [Github](https://github.com/WissamAntoun) | <wfa07@mail.aub.edu> | <wissam.antoun@gmail.com>

**Fady Baly**: [Linkedin](https://www.linkedin.com/in/fadybaly/) | [Twitter](https://twitter.com/fadybaly) | [Github](https://github.com/fadybaly) | <fgb06@mail.aub.edu> | <baly.fady@gmail.com>