Mihai-Dan MAŞALA (25095)
commited on
Commit
•
3b2670e
1
Parent(s):
23a9a61
Update README
Browse files
README.md
CHANGED
@@ -13,11 +13,11 @@ language:
|
|
13 |
Pretrained model on Romanian language using a masked language modeling (MLM) and next sentence prediction (NSP) objective.
|
14 |
It was introduced in this [paper](https://www.blank.org/). Three BERT models were released: **RoBERT-small**, RoBERT-base and RoBERT-large, all versions uncased.
|
15 |
|
16 |
-
| Model
|
17 |
-
|
18 |
-
| *RoBERT-small* | *19M*
|
19 |
-
| RoBERT-base
|
20 |
-
| RoBERT-large
|
21 |
|
22 |
|
23 |
|
@@ -53,12 +53,12 @@ outputs = model(**inputs)
|
|
53 |
|
54 |
The model is trained on the following compilation of corpora. Note that we present the statistics after the cleaning process.
|
55 |
|
56 |
-
Corpus
|
57 |
-
|
58 |
-
Oscar
|
59 |
-
RoTex
|
60 |
-
RoWiki
|
61 |
-
Total | 2.07B | 103M
|
62 |
|
63 |
|
64 |
## Eval results
|
@@ -67,45 +67,45 @@ Total | 2.07B | 103M | 12.6
|
|
67 |
|
68 |
We report Macro-averaged F1 score (in %)
|
69 |
|
70 |
-
Model
|
71 |
-
|
72 |
-
multilingual-BERT
|
73 |
-
XLM-R-base
|
74 |
-
BERT-base-ro
|
75 |
-
RoBERT-small
|
76 |
-
RoBERT-base
|
77 |
-
RoBERT-large
|
78 |
|
79 |
### Moldavian vs. Romanian Dialect and Cross-dialect Topic identification
|
80 |
|
81 |
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 %).
|
82 |
|
83 |
-
Model
|
84 |
-
|
85 |
-
2-CNN + SVM
|
86 |
-
Char+Word SVM
|
87 |
-
BiGRU
|
88 |
-
multilingual-BERT
|
89 |
-
XLM-R-base
|
90 |
-
BERT-base-ro
|
91 |
-
RoBERT-small
|
92 |
-
RoBERT-base
|
93 |
-
RoBERT-large
|
94 |
|
95 |
### Diacritics Restoration
|
96 |
|
97 |
Challenge can be found [here](https://diacritics-challenge.speed.pub.ro/). We report results on the official test set, as accuracies in %.
|
98 |
|
99 |
-
Model
|
100 |
-
|
101 |
-
BiLSTM
|
102 |
-
CharCNN
|
103 |
-
CharCNN + multilingual-BERT | 99.72
|
104 |
-
CharCNN + XLM-R-base
|
105 |
-
CharCNN + BERT-base-ro
|
106 |
-
CharCNN + RoBERT-small
|
107 |
-
CharCNN + RoBERT-base
|
108 |
-
CharCNN + RoBERT-large
|
109 |
|
110 |
|
111 |
### BibTeX entry and citation info
|
13 |
Pretrained model on Romanian language using a masked language modeling (MLM) and next sentence prediction (NSP) objective.
|
14 |
It was introduced in this [paper](https://www.blank.org/). Three BERT models were released: **RoBERT-small**, RoBERT-base and RoBERT-large, all versions uncased.
|
15 |
|
16 |
+
| Model | Weights | L | H | A | MLM accuracy | NSP accuracy |
|
17 |
+
|----------------|:---------:|:------:|:------:|:------:|:--------------:|:--------------:|
|
18 |
+
| *RoBERT-small* | *19M* | *12* | *256* | *8* | *0.5363* | *0.9687* |
|
19 |
+
| RoBERT-base | 114M | 12 | 768 | 12 | 0.6511 | 0.9802 |
|
20 |
+
| RoBERT-large | 341M | 24 | 1024 | 24 | 0.6929 | 0.9843 |
|
21 |
|
22 |
|
23 |
|
53 |
|
54 |
The model is trained on the following compilation of corpora. Note that we present the statistics after the cleaning process.
|
55 |
|
56 |
+
| Corpus | Words | Sentences | Size (GB)|
|
57 |
+
|-----------|-----------|-----------|----------|
|
58 |
+
| Oscar | 1.78B | 87M | 10.8 |
|
59 |
+
| RoTex | 240M | 14M | 1.5 |
|
60 |
+
| RoWiki | 50M | 2M | 0.3 |
|
61 |
+
| **Total** | **2.07B** | **103M** | **12.6** |
|
62 |
|
63 |
|
64 |
## Eval results
|
67 |
|
68 |
We report Macro-averaged F1 score (in %)
|
69 |
|
70 |
+
| Model | Dev | Test |
|
71 |
+
| -----------------|----------|----------|
|
72 |
+
| multilingual-BERT| 68.96 | 69.57 |
|
73 |
+
| XLM-R-base | 71.26 | 71.71 |
|
74 |
+
| BERT-base-ro | 70.49 | 71.02 |
|
75 |
+
| *RoBERT-small* | *66.32* | *66.37* |
|
76 |
+
| RoBERT-base | 70.89 | 71.61 |
|
77 |
+
| RoBERT-large | **72.48**| **72.11**|
|
78 |
|
79 |
### Moldavian vs. Romanian Dialect and Cross-dialect Topic identification
|
80 |
|
81 |
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 %).
|
82 |
|
83 |
+
| Model | Dialect Classification | MD to RO | RO to MD|
|
84 |
+
|-------------------|------------------------|----------|----------|
|
85 |
+
| 2-CNN + SVM | 93.40 | 65.09 | 75.21 |
|
86 |
+
| Char+Word SVM | 96.20 | 69.08 | 81.93 |
|
87 |
+
| BiGRU | 93.30 | **70.10**| 80.30 |
|
88 |
+
| multilingual-BERT | 95.34 | 68.76 | 78.24 |
|
89 |
+
| XLM-R-base | 96.28 | 69.93 | 82.28 |
|
90 |
+
| BERT-base-ro | 96.20 | 69.93 | 78.79 |
|
91 |
+
| *RoBERT-small* | *95.67* | *69.01* | *80.40* |
|
92 |
+
| RoBERT-base | 97.39 | 68.30 | 81.09 |
|
93 |
+
| RoBERT-large | **97.78** | 69.91 | **83.65**|
|
94 |
|
95 |
### Diacritics Restoration
|
96 |
|
97 |
Challenge can be found [here](https://diacritics-challenge.speed.pub.ro/). We report results on the official test set, as accuracies in %.
|
98 |
|
99 |
+
| Model | word level | char level |
|
100 |
+
|-----------------------------|------------|------------|
|
101 |
+
| BiLSTM | 99.42 | - |
|
102 |
+
| CharCNN | 98.40 | 99.65 |
|
103 |
+
| CharCNN + multilingual-BERT | 99.72 | 99.94 |
|
104 |
+
| CharCNN + XLM-R-base | 99.76 | **99.95** |
|
105 |
+
| CharCNN + BERT-base-ro | **99.79** | **99.95** |
|
106 |
+
| *CharCNN + RoBERT-small* | *99.73* | *99.94* |
|
107 |
+
| CharCNN + RoBERT-base | 99.78 | **99.95** |
|
108 |
+
| CharCNN + RoBERT-large | 99.76 | **99.95** |
|
109 |
|
110 |
|
111 |
### BibTeX entry and citation info
|