File size: 4,720 Bytes
09d62f9 5d01069 537489f 5d01069 b4a9f6b 537489f 60404f8 537489f 5d01069 537489f 5d01069 537489f 5d01069 537489f 5d01069 537489f 5d01069 cea027b bb7dc16 5d01069 cea027b 537489f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 |
Model card for RoBERT-base
---
language:
- ro
---
# RoBERT-base
## Pretrained BERT model for Romanian
Pretrained model on Romanian language using a masked language modeling (MLM) and next sentence prediction (NSP) objective.
It was introduced in this [paper](https://www.aclweb.org/anthology/2020.coling-main.581/). Three BERT models were released: RoBERT-small, **RoBERT-base** and RoBERT-large, all versions uncased.
| Model | Weights | L | H | A | MLM accuracy | NSP accuracy |
|----------------|:---------:|:------:|:------:|:------:|:--------------:|:--------------:|
| RoBERT-small | 19M | 12 | 256 | 8 | 0.5363 | 0.9687 |
| *RoBERT-base* | *114M* | *12* | *768* | *12* | *0.6511* | *0.9802* |
| RoBERT-large | 341M | 24 | 1024 | 24 | 0.6929 | 0.9843 |
All models are available:
* [RoBERT-small](https://huggingface.co/readerbench/RoBERT-small)
* [RoBERT-base](https://huggingface.co/readerbench/RoBERT-base)
* [RoBERT-large](https://huggingface.co/readerbench/RoBERT-large)
#### How to use
```python
# tensorflow
from transformers import AutoModel, AutoTokenizer, TFAutoModel
tokenizer = AutoTokenizer.from_pretrained("readerbench/RoBERT-base")
model = TFAutoModel.from_pretrained("readerbench/RoBERT-base")
inputs = tokenizer("exemplu de propoziție", return_tensors="tf")
outputs = model(inputs)
# pytorch
from transformers import AutoModel, AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("readerbench/RoBERT-base")
model = AutoModel.from_pretrained("readerbench/RoBERT-base")
inputs = tokenizer("exemplu de propoziție", return_tensors="pt")
outputs = model(**inputs)
```
## Training data
The model is trained on the following compilation of corpora. Note that we present the statistics after the cleaning process.
| Corpus | Words | Sentences | Size (GB)|
|-----------|:---------:|:---------:|:--------:|
| Oscar | 1.78B | 87M | 10.8 |
| RoTex | 240M | 14M | 1.5 |
| RoWiki | 50M | 2M | 0.3 |
| **Total** | **2.07B** | **103M** | **12.6** |
## Downstream performance
### Sentiment analysis
We report Macro-averaged F1 score (in %)
| Model | Dev | Test |
|------------------|:--------:|:--------:|
| multilingual-BERT| 68.96 | 69.57 |
| XLM-R-base | 71.26 | 71.71 |
| BERT-base-ro | 70.49 | 71.02 |
| RoBERT-small | 66.32 | 66.37 |
| *RoBERT-base* | *70.89* | *71.61* |
| RoBERT-large | **72.48**| **72.11**|
### Moldavian vs. Romanian Dialect and Cross-dialect Topic identification
We report results on [VarDial 2019](https://sites.google.com/view/vardial2019/campaign) Moldavian vs. Romanian Cross-dialect Topic identification Challenge, as Macro-averaged F1 score (in %).
| Model | Dialect Classification | MD to RO | RO to MD |
|-------------------|:----------------------:|:--------:|:--------:|
| 2-CNN + SVM | 93.40 | 65.09 | 75.21 |
| Char+Word SVM | 96.20 | 69.08 | 81.93 |
| BiGRU | 93.30 | **70.10**| 80.30 |
| multilingual-BERT | 95.34 | 68.76 | 78.24 |
| XLM-R-base | 96.28 | 69.93 | 82.28 |
| BERT-base-ro | 96.20 | 69.93 | 78.79 |
| RoBERT-small | 95.67 | 69.01 | 80.40 |
| *RoBERT-base* | *97.39* | *68.30* | *81.09* |
| RoBERT-large | **97.78** | 69.91 | **83.65**|
### Diacritics Restoration
Challenge can be found [here](https://diacritics-challenge.speed.pub.ro/). We report results on the official test set, as accuracies in %.
| Model | word level | char level |
|-----------------------------|:----------:|:----------:|
| BiLSTM | 99.42 | - |
| CharCNN | 98.40 | 99.65 |
| CharCNN + multilingual-BERT | 99.72 | 99.94 |
| CharCNN + XLM-R-base | 99.76 | **99.95** |
| CharCNN + BERT-base-ro | **99.79** | **99.95** |
| CharCNN + RoBERT-small | 99.73 | 99.94 |
| *CharCNN + RoBERT-base* | *99.78* | **99.95** |
| CharCNN + RoBERT-large | 99.76 | **99.95** |
### BibTeX entry and citation info
```bibtex
@inproceedings{masala2020robert,
title={RoBERT--A Romanian BERT Model},
author={Masala, Mihai and Ruseti, Stefan and Dascalu, Mihai},
booktitle={Proceedings of the 28th International Conference on Computational Linguistics},
pages={6626--6637},
year={2020}
}
```
|