Update README.md
Browse files
README.md
CHANGED
|
@@ -190,17 +190,6 @@ All of our models attained good accuracy values, in the range of 0.65, as can be
|
|
| 190 |
We are currently in the process of applying our language models to downstream tasks.
|
| 191 |
|
| 192 |
|
| 193 |
-
<figure>
|
| 194 |
-
|
| 195 |
-
<caption>Table x.</caption>
|
| 196 |
-
|
| 197 |
-
| Dataset | Metric | BERT-m | BERT-wwm | BSC-BNE | Beta | Random | Stepwise | Gaussian | Random-512 | Gaussian-512 |
|
| 198 |
-
|----------|---------------|--------|----------|----------|--------|---------|------------|-----------|--------------|---------------|
|
| 199 |
-
| CoNLL 2002-POS | F1 | 0.9629 | 0.9642 | 0.9659 | 0.9638 | 0.9656 | 0.9656 | **0.9662** | 0.9660 | **0.9662** |
|
| 200 |
-
| CoNLL 2002-POS | F1 | 0.9687 | 0.9700 | 0.9707 | 0.9690 | 0.9704 | 0.9707 | 0.9709 | 0.9707 | **0.9714** |
|
| 201 |
-
|
| 202 |
-
</figure>
|
| 203 |
-
|
| 204 |
<figure>
|
| 205 |
|
| 206 |
<caption>Table x. Dataset for POS nad NER is CoNLL 2002.</caption>
|
|
@@ -224,50 +213,12 @@ We are currently in the process of applying our language models to downstream ta
|
|
| 224 |
Using sequence length 128 we have achieved exact match 50.96 and F1 68.74.
|
| 225 |
|
| 226 |
|
| 227 |
-
|
| 228 |
All models trained with max length 512 and batch size 8, using the CoNLL 2002 dataset.
|
| 229 |
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
<caption>Table 3. Results for POS.</caption>
|
| 233 |
-
|
| 234 |
-
| Model | F1 | Accuracy |
|
| 235 |
-
|----------------------------------------------------|----------|----------|
|
| 236 |
-
| bert-base-multilingual-cased | 0.9629 | 0.9687 |
|
| 237 |
-
| dccuchile/bert-base-spanish-wwm-cased | 0.9642 | 0.9700 |
|
| 238 |
-
| BSC-TeMU/roberta-base-bne | 0.9659 | 0.9707 |
|
| 239 |
-
| bertin-project/bertin-roberta-base-spanish | 0.9638 | 0.9690 |
|
| 240 |
-
| bertin-project/bertin-base-random | 0.9656 | 0.9704 |
|
| 241 |
-
| bertin-project/bertin-base-stepwise | 0.9656 | 0.9707 |
|
| 242 |
-
| bertin-project/bertin-base-gaussian | **0.9662** | 0.9709 |
|
| 243 |
-
| bertin-project/bertin-base-random-exp-512seqlen | 0.9660 | 0.9707 |
|
| 244 |
-
| bertin-project/bertin-base-gaussian-exp-512seqlen | **0.9662** | **0.9714** |
|
| 245 |
-
|
| 246 |
-
</figure>
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
## NER
|
| 250 |
All models trained with max length 512 and batch size 8, using the CoNLL 2002 dataset.
|
| 251 |
|
| 252 |
-
<figure>
|
| 253 |
-
|
| 254 |
-
<caption>Table 4. Results for NER.</caption>
|
| 255 |
-
|
| 256 |
-
| Model | F1 | Accuracy |
|
| 257 |
-
|----------------------------------------------------|----------|----------|
|
| 258 |
-
| bert-base-multilingual-cased | 0.8539 | 0.9779 |
|
| 259 |
-
| dccuchile/bert-base-spanish-wwm-cased | 0.8579 | 0.9783 |
|
| 260 |
-
| BSC-TeMU/roberta-base-bne | 0.8700 | 0.9807 |
|
| 261 |
-
| bertin-project/bertin-roberta-base-spanish | 0.8725 | 0.9812 |
|
| 262 |
-
| bertin-project/bertin-base-random | 0.8704 | 0.9807 |
|
| 263 |
-
| bertin-project/bertin-base-stepwise | 0.8705 | 0.9809 |
|
| 264 |
-
| bertin-project/bertin-base-gaussian | **0.8792** | **0.9816** |
|
| 265 |
-
| bertin-project/bertin-base-random-exp-512seqlen | 0.8616 | 0.9803 |
|
| 266 |
-
| bertin-project/bertin-base-gaussian-exp-512seqlen | **0.8764** | **0.9819** |
|
| 267 |
-
|
| 268 |
-
</figure>
|
| 269 |
-
|
| 270 |
-
|
| 271 |
## PAWS-X
|
| 272 |
All models trained with max length 512 and batch size 8. These numbers are surprising both for the repeated instances of 0.5765 accuracy and for the large differences in performance. However, experiments have been repeated several times and the results are consistent.
|
| 273 |
|
| 190 |
We are currently in the process of applying our language models to downstream tasks.
|
| 191 |
|
| 192 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
<figure>
|
| 194 |
|
| 195 |
<caption>Table x. Dataset for POS nad NER is CoNLL 2002.</caption>
|
| 213 |
Using sequence length 128 we have achieved exact match 50.96 and F1 68.74.
|
| 214 |
|
| 215 |
|
| 216 |
+
POS
|
| 217 |
All models trained with max length 512 and batch size 8, using the CoNLL 2002 dataset.
|
| 218 |
|
| 219 |
+
NER
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 220 |
All models trained with max length 512 and batch size 8, using the CoNLL 2002 dataset.
|
| 221 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 222 |
## PAWS-X
|
| 223 |
All models trained with max length 512 and batch size 8. These numbers are surprising both for the repeated instances of 0.5765 accuracy and for the large differences in performance. However, experiments have been repeated several times and the results are consistent.
|
| 224 |
|