Commit
•
c96c930
1
Parent(s):
38ea57a
Update README.md
Browse files
README.md
CHANGED
|
@@ -173,8 +173,30 @@ This is also a base-sized model with many languages and many tasks, so capacity
|
|
| 173 |
|
| 174 |
# Evaluation
|
| 175 |
In these metrics, keep in mind that
|
| 176 |
-
1.
|
| 177 |
-
2. Sentence boundaries and true-casing
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 178 |
|
| 179 |
<details>
|
| 180 |
<summary>English</summary>
|
|
@@ -182,10 +204,10 @@ In these metrics, keep in mind that
|
|
| 182 |
```
|
| 183 |
punct_post test report:
|
| 184 |
label precision recall f1 support
|
| 185 |
-
<NULL> (label_id: 0) 98.71 98.
|
| 186 |
-
. (label_id: 1) 87.
|
| 187 |
-
, (label_id: 2)
|
| 188 |
-
? (label_id: 3)
|
| 189 |
？ (label_id: 4) 0.00 0.00 0.00 0
|
| 190 |
， (label_id: 5) 0.00 0.00 0.00 0
|
| 191 |
。 (label_id: 6) 0.00 0.00 0.00 0
|
|
@@ -199,26 +221,225 @@ punct_post test report:
|
|
| 199 |
፣ (label_id: 14) 0.00 0.00 0.00 0
|
| 200 |
፧ (label_id: 15) 0.00 0.00 0.00 0
|
| 201 |
-------------------
|
| 202 |
-
micro avg 97.
|
| 203 |
-
macro avg 83.
|
| 204 |
-
weighted avg 97.
|
| 205 |
|
| 206 |
cap test report:
|
| 207 |
label precision recall f1 support
|
| 208 |
-
LOWER (label_id: 0) 99.
|
| 209 |
-
UPPER (label_id: 1) 89.
|
| 210 |
-------------------
|
| 211 |
-
micro avg 99.15 99.15 99.15
|
| 212 |
-
macro avg 94.
|
| 213 |
-
weighted avg 99.17 99.15 99.16
|
| 214 |
|
| 215 |
seg test report:
|
| 216 |
label precision recall f1 support
|
| 217 |
-
NOSTOP (label_id: 0) 99.
|
| 218 |
-
FULLSTOP (label_id: 1) 89.
|
| 219 |
-------------------
|
| 220 |
-
micro avg 98.
|
| 221 |
-
macro avg 94.
|
| 222 |
-
weighted avg 98.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 224 |
</details>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 173 |
|
| 174 |
# Evaluation
|
| 175 |
In these metrics, keep in mind that
|
| 176 |
+
1. The data is noisy
|
| 177 |
+
2. Sentence boundaries and true-casing are conditioned on predicted punctuation, which is the most difficult task and sometimes incorrect.
|
| 178 |
+
When conditioning on reference punctuation, true-casing and SBD is practically 100% for most languages.
|
| 179 |
+
4. Punctuation can be subjective. E.g.,
|
| 180 |
+
|
| 181 |
+
`Hola mundo, ¿cómo estás?`
|
| 182 |
+
|
| 183 |
+
or
|
| 184 |
+
|
| 185 |
+
`Hola mundo. ¿Cómo estás?`
|
| 186 |
+
|
| 187 |
+
When the sentences are longer and more practical, these ambiguities abound and affect all 3 analytics.
|
| 188 |
+
|
| 189 |
+
|
| 190 |
+
## Selected Language Evaluation Reports
|
| 191 |
+
Each test example was generated using the following procedure:
|
| 192 |
+
|
| 193 |
+
1. Concatenate 5 random sentences
|
| 194 |
+
2. Lower-case the concatenated sentence
|
| 195 |
+
3. Remove all punctuation
|
| 196 |
+
|
| 197 |
+
The data is a held-out portion of News Crawl, which has been deduplicated.
|
| 198 |
+
2,000 lines of data per language was used, generating 2,000 unique examples of 5 sentences each.
|
| 199 |
+
The last 4 sentences of each example were randomly sampled from the 2,000 and may be duplicated.
|
| 200 |
|
| 201 |
<details>
|
| 202 |
<summary>English</summary>
|
|
|
|
| 204 |
```
|
| 205 |
punct_post test report:
|
| 206 |
label precision recall f1 support
|
| 207 |
+
<NULL> (label_id: 0) 98.71 98.66 98.68 156605
|
| 208 |
+
. (label_id: 1) 87.72 88.85 88.28 8752
|
| 209 |
+
, (label_id: 2) 68.06 67.81 67.93 5216
|
| 210 |
+
? (label_id: 3) 79.38 77.20 78.27 693
|
| 211 |
？ (label_id: 4) 0.00 0.00 0.00 0
|
| 212 |
， (label_id: 5) 0.00 0.00 0.00 0
|
| 213 |
。 (label_id: 6) 0.00 0.00 0.00 0
|
|
|
|
| 221 |
፣ (label_id: 14) 0.00 0.00 0.00 0
|
| 222 |
፧ (label_id: 15) 0.00 0.00 0.00 0
|
| 223 |
-------------------
|
| 224 |
+
micro avg 97.13 97.13 97.13 171266
|
| 225 |
+
macro avg 83.46 83.13 83.29 171266
|
| 226 |
+
weighted avg 97.13 97.13 97.13 171266
|
| 227 |
|
| 228 |
cap test report:
|
| 229 |
label precision recall f1 support
|
| 230 |
+
LOWER (label_id: 0) 99.63 99.49 99.56 526612
|
| 231 |
+
UPPER (label_id: 1) 89.19 91.84 90.50 24161
|
| 232 |
-------------------
|
| 233 |
+
micro avg 99.15 99.15 99.15 550773
|
| 234 |
+
macro avg 94.41 95.66 95.03 550773
|
| 235 |
+
weighted avg 99.17 99.15 99.16 550773
|
| 236 |
|
| 237 |
seg test report:
|
| 238 |
label precision recall f1 support
|
| 239 |
+
NOSTOP (label_id: 0) 99.37 99.42 99.39 162044
|
| 240 |
+
FULLSTOP (label_id: 1) 89.75 88.84 89.29 9222
|
| 241 |
-------------------
|
| 242 |
+
micro avg 98.85 98.85 98.85 171266
|
| 243 |
+
macro avg 94.56 94.13 94.34 171266
|
| 244 |
+
weighted avg 98.85 98.85 98.85 171266
|
| 245 |
+
```
|
| 246 |
+
</details>
|
| 247 |
+
|
| 248 |
+
|
| 249 |
+
<details>
|
| 250 |
+
<summary>Spanish</summary>
|
| 251 |
+
|
| 252 |
```
|
| 253 |
+
punct_pre test report:
|
| 254 |
+
label precision recall f1 support
|
| 255 |
+
<NULL> (label_id: 0) 99.94 99.92 99.93 185535
|
| 256 |
+
¿ (label_id: 1) 55.01 64.86 59.53 296
|
| 257 |
+
-------------------
|
| 258 |
+
micro avg 99.86 99.86 99.86 185831
|
| 259 |
+
macro avg 77.48 82.39 79.73 185831
|
| 260 |
+
weighted avg 99.87 99.86 99.87 185831
|
| 261 |
+
|
| 262 |
+
punct_post test report:
|
| 263 |
+
label precision recall f1 support
|
| 264 |
+
<NULL> (label_id: 0) 98.74 98.86 98.80 170282
|
| 265 |
+
. (label_id: 1) 90.07 89.58 89.82 9959
|
| 266 |
+
, (label_id: 2) 68.33 67.00 67.66 5300
|
| 267 |
+
? (label_id: 3) 70.25 58.62 63.91 290
|
| 268 |
+
？ (label_id: 4) 0.00 0.00 0.00 0
|
| 269 |
+
， (label_id: 5) 0.00 0.00 0.00 0
|
| 270 |
+
。 (label_id: 6) 0.00 0.00 0.00 0
|
| 271 |
+
、 (label_id: 7) 0.00 0.00 0.00 0
|
| 272 |
+
・ (label_id: 8) 0.00 0.00 0.00 0
|
| 273 |
+
। (label_id: 9) 0.00 0.00 0.00 0
|
| 274 |
+
؟ (label_id: 10) 0.00 0.00 0.00 0
|
| 275 |
+
، (label_id: 11) 0.00 0.00 0.00 0
|
| 276 |
+
; (label_id: 12) 0.00 0.00 0.00 0
|
| 277 |
+
። (label_id: 13) 0.00 0.00 0.00 0
|
| 278 |
+
፣ (label_id: 14) 0.00 0.00 0.00 0
|
| 279 |
+
፧ (label_id: 15) 0.00 0.00 0.00 0
|
| 280 |
+
-------------------
|
| 281 |
+
micro avg 97.39 97.39 97.39 185831
|
| 282 |
+
macro avg 81.84 78.51 80.05 185831
|
| 283 |
+
weighted avg 97.36 97.39 97.37 185831
|
| 284 |
+
|
| 285 |
+
cap test report:
|
| 286 |
+
label precision recall f1 support
|
| 287 |
+
LOWER (label_id: 0) 99.62 99.60 99.61 555041
|
| 288 |
+
UPPER (label_id: 1) 90.60 91.06 90.83 23538
|
| 289 |
+
-------------------
|
| 290 |
+
micro avg 99.25 99.25 99.25 578579
|
| 291 |
+
macro avg 95.11 95.33 95.22 578579
|
| 292 |
+
weighted avg 99.25 99.25 99.25 578579
|
| 293 |
+
|
| 294 |
+
[NeMo I 2023-02-22 17:24:04 punct_cap_seg_model:427] seg test report:
|
| 295 |
+
label precision recall f1 support
|
| 296 |
+
NOSTOP (label_id: 0) 99.44 99.54 99.49 175908
|
| 297 |
+
FULLSTOP (label_id: 1) 91.68 89.98 90.82 9923
|
| 298 |
+
-------------------
|
| 299 |
+
micro avg 99.03 99.03 99.03 185831
|
| 300 |
+
macro avg 95.56 94.76 95.16 185831
|
| 301 |
+
weighted avg 99.02 99.03 99.02 185831
|
| 302 |
+
```
|
| 303 |
</details>
|
| 304 |
+
|
| 305 |
+
<details>
|
| 306 |
+
<summary>Chinese</summary>
|
| 307 |
+
|
| 308 |
+
```
|
| 309 |
+
punct_post test report:
|
| 310 |
+
label precision recall f1 support
|
| 311 |
+
<NULL> (label_id: 0) 98.82 97.34 98.07 147920
|
| 312 |
+
. (label_id: 1) 0.00 0.00 0.00 0
|
| 313 |
+
, (label_id: 2) 0.00 0.00 0.00 0
|
| 314 |
+
? (label_id: 3) 0.00 0.00 0.00 0
|
| 315 |
+
？ (label_id: 4) 85.77 80.71 83.16 560
|
| 316 |
+
， (label_id: 5) 59.88 78.02 67.75 6901
|
| 317 |
+
。 (label_id: 6) 92.50 93.92 93.20 10988
|
| 318 |
+
、 (label_id: 7) 0.00 0.00 0.00 0
|
| 319 |
+
・ (label_id: 8) 0.00 0.00 0.00 0
|
| 320 |
+
। (label_id: 9) 0.00 0.00 0.00 0
|
| 321 |
+
؟ (label_id: 10) 0.00 0.00 0.00 0
|
| 322 |
+
، (label_id: 11) 0.00 0.00 0.00 0
|
| 323 |
+
; (label_id: 12) 0.00 0.00 0.00 0
|
| 324 |
+
። (label_id: 13) 0.00 0.00 0.00 0
|
| 325 |
+
፣ (label_id: 14) 0.00 0.00 0.00 0
|
| 326 |
+
፧ (label_id: 15) 0.00 0.00 0.00 0
|
| 327 |
+
-------------------
|
| 328 |
+
micro avg 96.25 96.25 96.25 166369
|
| 329 |
+
macro avg 84.24 87.50 85.55 166369
|
| 330 |
+
weighted avg 96.75 96.25 96.45 166369
|
| 331 |
+
|
| 332 |
+
cap test report:
|
| 333 |
+
label precision recall f1 support
|
| 334 |
+
LOWER (label_id: 0) 97.07 92.39 94.67 394
|
| 335 |
+
UPPER (label_id: 1) 70.59 86.75 77.84 83
|
| 336 |
+
-------------------
|
| 337 |
+
micro avg 91.40 91.40 91.40 477
|
| 338 |
+
macro avg 83.83 89.57 86.25 477
|
| 339 |
+
weighted avg 92.46 91.40 91.74 477
|
| 340 |
+
|
| 341 |
+
seg test report:
|
| 342 |
+
label precision recall f1 support
|
| 343 |
+
NOSTOP (label_id: 0) 99.58 99.53 99.56 156369
|
| 344 |
+
FULLSTOP (label_id: 1) 92.77 93.50 93.13 10000
|
| 345 |
+
-------------------
|
| 346 |
+
micro avg 99.17 99.17 99.17 166369
|
| 347 |
+
macro avg 96.18 96.52 96.35 166369
|
| 348 |
+
weighted avg 99.17 99.17 99.17 166369
|
| 349 |
+
```
|
| 350 |
+
</details>
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
<details>
|
| 354 |
+
<summary>Hindi</summary>
|
| 355 |
+
|
| 356 |
+
```
|
| 357 |
+
punct_post test report:
|
| 358 |
+
label precision recall f1 support
|
| 359 |
+
<NULL> (label_id: 0) 99.58 99.59 99.59 176743
|
| 360 |
+
. (label_id: 1) 0.00 0.00 0.00 0
|
| 361 |
+
, (label_id: 2) 68.32 65.23 66.74 1815
|
| 362 |
+
? (label_id: 3) 60.27 44.90 51.46 98
|
| 363 |
+
？ (label_id: 4) 0.00 0.00 0.00 0
|
| 364 |
+
， (label_id: 5) 0.00 0.00 0.00 0
|
| 365 |
+
。 (label_id: 6) 0.00 0.00 0.00 0
|
| 366 |
+
、 (label_id: 7) 0.00 0.00 0.00 0
|
| 367 |
+
・ (label_id: 8) 0.00 0.00 0.00 0
|
| 368 |
+
। (label_id: 9) 96.45 97.43 96.94 10136
|
| 369 |
+
؟ (label_id: 10) 0.00 0.00 0.00 0
|
| 370 |
+
، (label_id: 11) 0.00 0.00 0.00 0
|
| 371 |
+
; (label_id: 12) 0.00 0.00 0.00 0
|
| 372 |
+
። (label_id: 13) 0.00 0.00 0.00 0
|
| 373 |
+
፣ (label_id: 14) 0.00 0.00 0.00 0
|
| 374 |
+
፧ (label_id: 15) 0.00 0.00 0.00 0
|
| 375 |
+
-------------------
|
| 376 |
+
micro avg 99.11 99.11 99.11 188792
|
| 377 |
+
macro avg 81.16 76.79 78.68 188792
|
| 378 |
+
weighted avg 99.10 99.11 99.10 188792
|
| 379 |
+
|
| 380 |
+
cap test report:
|
| 381 |
+
label precision recall f1 support
|
| 382 |
+
LOWER (label_id: 0) 98.25 95.06 96.63 708
|
| 383 |
+
UPPER (label_id: 1) 89.46 96.12 92.67 309
|
| 384 |
+
-------------------
|
| 385 |
+
micro avg 95.38 95.38 95.38 1017
|
| 386 |
+
macro avg 93.85 95.59 94.65 1017
|
| 387 |
+
weighted avg 95.58 95.38 95.42 1017
|
| 388 |
+
|
| 389 |
+
seg test report:
|
| 390 |
+
label precision recall f1 support
|
| 391 |
+
NOSTOP (label_id: 0) 99.87 99.85 99.86 178892
|
| 392 |
+
FULLSTOP (label_id: 1) 97.38 97.58 97.48 9900
|
| 393 |
+
-------------------
|
| 394 |
+
micro avg 99.74 99.74 99.74 188792
|
| 395 |
+
macro avg 98.62 98.72 98.67 188792
|
| 396 |
+
weighted avg 99.74 99.74 99.74 188792
|
| 397 |
+
```
|
| 398 |
+
</details>
|
| 399 |
+
|
| 400 |
+
<details>
|
| 401 |
+
<summary>Amharic</summary>
|
| 402 |
+
|
| 403 |
+
```
|
| 404 |
+
punct_post test report:
|
| 405 |
+
label precision recall f1 support
|
| 406 |
+
<NULL> (label_id: 0) 99.58 99.42 99.50 236298
|
| 407 |
+
. (label_id: 1) 0.00 0.00 0.00 0
|
| 408 |
+
, (label_id: 2) 0.00 0.00 0.00 0
|
| 409 |
+
? (label_id: 3) 0.00 0.00 0.00 0
|
| 410 |
+
？ (label_id: 4) 0.00 0.00 0.00 0
|
| 411 |
+
， (label_id: 5) 0.00 0.00 0.00 0
|
| 412 |
+
。 (label_id: 6) 0.00 0.00 0.00 0
|
| 413 |
+
、 (label_id: 7) 0.00 0.00 0.00 0
|
| 414 |
+
・ (label_id: 8) 0.00 0.00 0.00 0
|
| 415 |
+
। (label_id: 9) 0.00 0.00 0.00 0
|
| 416 |
+
؟ (label_id: 10) 0.00 0.00 0.00 0
|
| 417 |
+
، (label_id: 11) 0.00 0.00 0.00 0
|
| 418 |
+
; (label_id: 12) 0.00 0.00 0.00 0
|
| 419 |
+
። (label_id: 13) 89.79 95.24 92.44 9169
|
| 420 |
+
፣ (label_id: 14) 66.85 56.58 61.29 1504
|
| 421 |
+
፧ (label_id: 15) 67.67 83.72 74.84 215
|
| 422 |
+
-------------------
|
| 423 |
+
micro avg 98.99 98.99 98.99 247186
|
| 424 |
+
macro avg 80.97 83.74 82.02 247186
|
| 425 |
+
weighted avg 98.99 98.99 98.98 247186
|
| 426 |
+
|
| 427 |
+
cap test report:
|
| 428 |
+
label precision recall f1 support
|
| 429 |
+
LOWER (label_id: 0) 96.65 99.78 98.19 1360
|
| 430 |
+
UPPER (label_id: 1) 98.90 85.13 91.50 316
|
| 431 |
+
-------------------
|
| 432 |
+
micro avg 97.02 97.02 97.02 1676
|
| 433 |
+
macro avg 97.77 92.45 94.84 1676
|
| 434 |
+
weighted avg 97.08 97.02 96.93 1676
|
| 435 |
+
|
| 436 |
+
seg test report:
|
| 437 |
+
label precision recall f1 support
|
| 438 |
+
NOSTOP (label_id: 0) 99.85 99.74 99.80 239845
|
| 439 |
+
FULLSTOP (label_id: 1) 91.72 95.25 93.45 7341
|
| 440 |
+
-------------------
|
| 441 |
+
micro avg 99.60 99.60 99.60 247186
|
| 442 |
+
macro avg 95.79 97.49 96.62 247186
|
| 443 |
+
weighted avg 99.61 99.60 99.61 247186
|
| 444 |
+
```
|
| 445 |
+
</details>
|