Closed buriy closed 4 years ago
Vocab: https://gist.github.com/sskorol/3f654f257c57f04fec4a6244402d2ae8 width = 150
Itn Tag Loss Tag % Dep Loss UAS LAS NER Loss NER P NER R NER F Token % CPU WPS GPU WPS
--- --------- -------- --------- ------ ------ --------- ------ ------ ------ ------- ------- -------
1 195670.409 89.856 672265.634 85.334 79.454 0.000 0.000 0.000 0.000 100.000 5470 10082
2 123066.648 91.355 519681.819 87.293 82.405 0.000 0.000 0.000 0.000 100.000 5470 10164
3 106329.392 92.066 466157.118 88.241 83.827 0.000 0.000 0.000 0.000 100.000 5470 10265
4 95541.371 92.501 431755.374 88.849 84.679 0.000 0.000 0.000 0.000 100.000 5470 10120
5 87373.762 92.830 405544.697 89.302 85.328 0.000 0.000 0.000 0.000 100.000 5470 10165
6 81018.798 93.035 386272.139 89.593 85.777 0.000 0.000 0.000 0.000 100.000 5470 10161
7 76154.969 93.221 369641.625 89.955 86.197 0.000 0.000 0.000 0.000 100.000 5470 10254
8 71413.536 93.303 356122.455 90.107 86.369 0.000 0.000 0.000 0.000 100.000 5470 10277
9 68078.113 93.377 342631.283 90.176 86.483 0.000 0.000 0.000 0.000 100.000 5470 10199
10 64366.264 93.459 333533.511 90.230 86.599 0.000 0.000 0.000 0.000 100.000 5470 10203
11 61935.280 93.583 323958.284 90.339 86.788 0.000 0.000 0.000 0.000 100.000 5470 10204
12 59330.765 93.604 315470.553 90.440 86.907 0.000 0.000 0.000 0.000 100.000 5470 10233
13 57306.374 93.643 306508.847 90.552 87.035 0.000 0.000 0.000 0.000 100.000 5470 10159
14 55086.122 93.636 300465.398 90.598 87.090 0.000 0.000 0.000 0.000 100.000 5470 10084
15 53226.834 93.667 293474.921 90.596 87.139 0.000 0.000 0.000 0.000 100.000 5470 9950
16 51378.742 93.736 286699.144 90.597 87.164 0.000 0.000 0.000 0.000 100.000 5470 9894
17 49850.338 93.753 281217.184 90.558 87.154 0.000 0.000 0.000 0.000 100.000 5470 10332
18 48744.009 93.771 275373.145 90.627 87.284 0.000 0.000 0.000 0.000 100.000 5470 10339
19 47157.016 93.761 270591.588 90.720 87.332 0.000 0.000 0.000 0.000 100.000 5470 10324
20 45887.111 93.772 267271.456 90.718 87.338 0.000 0.000 0.000 0.000 100.000 5470 10285
vec = norms_from_navec, width=96, https://gist.github.com/buriy/f81330ccd5f35e503f957e96844540e3/revisions#diff-27299c9d2f12875b2e5f4fce3ce484f3:
Itn Tag Loss Tag % Dep Loss UAS LAS NER L
--- --------- -------- --------- ------ ------ -----
1 217942.606 88.781 716078.814 83.983 77.543
2 141262.220 90.420 572827.089 86.235 80.805
3 124264.333 91.266 518904.142 87.323 82.419
4 113914.387 91.712 482734.570 87.997 83.472
5 106446.137 92.102 458973.689 88.457 84.146
6 100019.644 92.343 439558.357 88.811 84.642
7 95041.872 92.522 422688.017 89.092 85.050
8 91603.579 92.685 409935.470 89.272 85.307
9 87856.239 92.822 399263.286 89.391 85.522
10 85022.764 92.951 387969.582 89.481 85.629
11 82410.319 93.015 379647.275 89.584 85.783
12 80401.838 93.084 372475.272 89.784 86.044
13 78053.943 93.168 366079.551 89.851 86.145
14 76207.141 93.235 360122.291 89.884 86.201
15 74474.127 93.297 354038.745 89.948 86.304
16 72706.680 93.358 347340.824 90.061 86.456
17 71683.732 93.386 341853.629 90.116 86.537
18 70011.172 93.439 340336.212 90.126 86.556
19 69002.519 93.447 334133.481 90.228 86.665
20 68094.355 93.488 330321.146 90.332 86.811
21 66957.630 93.512 327471.405 90.365 86.871
22 65769.578 93.515 321939.197 90.379 86.890
23 64771.484 93.563 320606.388 90.357 86.905
24 64398.250 93.585 317066.967 90.370 86.915
25 63262.954 93.593 314013.264 90.433 86.987
26 62345.851 93.628 310481.751 90.451 87.005
27 61370.863 93.637 308041.837 90.468 87.049
28 61462.744 93.623 307849.631 90.498 87.070
29 60199.385 93.649 301794.233 90.494 87.061
30 59873.190 93.661 301905.787 90.528 87.101
No vectors, width=150
Itn Tag Loss Tag % Dep Loss UAS LAS
--- --------- -------- --------- ------ ------
1 200550.240 89.650 683498.798 85.063 79.134
2 127058.552 91.217 528910.494 87.142 82.170
3 110075.959 91.884 473528.988 88.208 83.618
4 99022.254 92.303 437539.379 88.860 84.571
5 91371.813 92.559 409869.130 89.174 85.010
6 84689.915 92.734 390382.618 89.518 85.538
7 80111.442 92.938 376035.951 89.641 85.795
8 75410.372 93.085 362495.646 89.886 86.105
9 72101.657 93.129 350447.011 90.066 86.322
10 68761.490 93.236 337636.480 90.223 86.539
11 66325.080 93.309 328696.397 90.276 86.642
12 64174.671 93.360 320164.963 90.243 86.676
13 61822.824 93.413 312061.878 90.382 86.857
14 59674.548 93.442 304749.814 90.498 87.021
15 57705.098 93.528 300285.224 90.531 87.044
16 55977.893 93.561 293359.634 90.627 87.202
17 54286.165 93.547 287182.622 90.592 87.181
18 53102.329 93.563 281234.365 90.609 87.233
19 51901.250 93.573 278506.581 90.643 87.273
20 51107.783 93.586 273447.448 90.670 87.316
21 49394.585 93.599 268765.360 90.770 87.409
22 48609.565 93.614 265592.125 90.736 87.372
23 47323.302 93.622 259167.853 90.815 87.435
24 46299.187 93.663 257274.539 90.789 87.389
25 45238.024 93.688 253004.762 90.793 87.427
26 44462.939 93.678 249416.752 90.738 87.384
27 43852.407 93.675 246350.867 90.763 87.432
28 43061.078 93.705 244398.172 90.728 87.393
29 42539.750 93.682 241708.512 90.769 87.476
30 41564.851 93.683 239089.832 90.814 87.563
vec = norms_from_navec, width=150, https://gist.github.com/buriy/f81330ccd5f35e503f957e96844540e3/revisions#diff-27299c9d2f12875b2e5f4fce3ce484f3:
1 196810.935 89.901 674082.843 85.317 79.467
2 123477.646 91.380 523418.743 87.304 82.415
3 106914.275 92.034 469262.359 88.214 83.830
4 95881.340 92.474 435113.845 88.763 84.616
5 87536.242 92.732 407065.647 89.154 85.151
6 81377.265 92.992 388305.329 89.533 85.629
7 76279.151 93.135 372260.317 89.724 85.935
8 71530.780 93.297 356597.583 90.060 86.362
9 68102.305 93.367 343777.439 90.145 86.477
10 64555.731 93.440 335099.408 90.244 86.667
11 61823.762 93.514 323844.126 90.359 86.827
12 59382.968 93.578 315458.675 90.385 86.875
13 57174.704 93.619 306645.989 90.409 86.928
14 55390.045 93.703 298922.144 90.473 86.986
15 53294.908 93.741 294415.630 90.562 87.131
16 51505.326 93.728 286795.211 90.550 87.144
17 49953.103 93.787 281430.921 90.614 87.233
18 48327.442 93.778 277675.906 90.652 87.295
19 47433.790 93.814 271525.058 90.681 87.320
20 46165.790 93.848 267326.015 90.710 87.362
21 45132.669 93.847 262522.712 90.757 87.406
22 43535.840 93.851 258846.360 90.841 87.510
23 42617.882 93.875 253999.910 90.820 87.519
24 42065.062 93.890 250185.802 90.879 87.581
25 40989.635 93.909 246938.856 90.910 87.618
26 39756.463 93.891 243510.403 90.894 87.637
27 39459.779 93.907 240215.944 90.882 87.605
28 38455.344 93.902 237592.473 90.860 87.559
29 37653.153 93.924 234728.028 90.919 87.633
30 37382.722 93.913 230929.532 90.936 87.666
Same args, different preparation script: https://gist.github.com/sskorol/2dcc110a58e932810ec55671e849a979
Itn Tag Loss Tag % Dep Loss UAS LAS
--- --------- -------- --------- ------ ------
1 200112.873 89.756 675646.251 85.087 79.200
2 124427.645 91.356 522325.208 87.304 82.381
3 107209.685 92.043 467316.867 88.067 83.612
4 96644.060 92.461 431185.378 88.596 84.386
5 88648.057 92.815 404570.997 89.096 85.078
6 81967.809 92.944 384839.287 89.466 85.587
7 77161.391 93.122 366746.242 89.776 85.998
8 72891.981 93.242 353431.963 89.968 86.227
9 69274.212 93.326 342155.651 90.045 86.356
10 65959.241 93.384 330291.187 90.184 86.503
11 63091.543 93.458 322220.869 90.170 86.585
12 61104.035 93.528 311960.291 90.257 86.682
13 58465.460 93.581 304510.071 90.311 86.768
14 56823.556 93.638 297195.149 90.422 86.921
15 55011.108 93.693 290235.511 90.503 87.053
16 52995.132 93.709 284636.749 90.481 87.028
17 51415.107 93.739 279206.300 90.620 87.188
18 49913.032 93.730 273005.355 90.620 87.253
19 48875.219 93.706 267365.928 90.675 87.331
20 47712.145 93.744 263814.256 90.670 87.311
21 46486.756 93.743 261534.572 90.689 87.352
22 45320.962 93.776 255665.435 90.777 87.459
23 44482.011 93.801 252113.816 90.768 87.451
24 43456.462 93.786 248063.270 90.744 87.444
25 42495.427 93.787 245076.256 90.744 87.485
26 41839.088 93.810 241856.637 90.721 87.464
27 40520.905 93.824 238334.513 90.724 87.483
28 40194.237 93.805 234280.121 90.797 87.564
29 39270.977 93.829 231577.732 90.776 87.547
30 38981.384 93.827 229523.775 90.861 87.624
Fixed script: https://gist.github.com/sskorol/2dcc110a58e932810ec55671e849a979
Itn Tag Loss Tag % Dep Loss UAS LAS
--- --------- -------- --------- ------ ------
1 161145.583 92.742 632702.367 86.757 81.768
2 93855.727 93.684 479985.515 88.517 84.389
3 80398.103 94.214 429999.728 89.312 85.553
4 72063.494 94.456 396952.023 89.880 86.432
5 66128.879 94.642 373615.299 90.050 86.731
6 61062.543 94.739 353606.956 90.396 87.187
7 57032.251 94.825 338820.866 90.516 87.351
8 53676.331 94.913 324472.700 90.627 87.539
9 50491.066 94.948 313799.272 90.769 87.725
10 47884.488 94.972 301719.675 90.881 87.877
11 45713.239 95.014 294273.747 90.966 87.990
12 43999.539 95.054 284167.030 90.994 88.088
13 42280.597 95.092 278617.027 91.045 88.168
14 40497.084 95.091 271569.468 91.064 88.239
15 38970.900 95.058 265302.393 91.088 88.282
16 37319.622 95.052 259257.846 91.173 88.372
17 36727.700 95.080 254876.380 91.216 88.412
18 35479.840 95.076 249545.049 91.250 88.460
19 34442.474 95.083 244911.117 91.299 88.524
20 33085.346 95.082 239358.418 91.304 88.526
21 32468.793 95.114 235457.510 91.376 88.576
22 31371.001 95.129 232571.953 91.386 88.591
23 30669.177 95.124 228429.651 91.382 88.597
24 30256.596 95.123 226056.217 91.336 88.542
25 29340.200 95.140 221773.170 91.348 88.589
26 28469.800 95.152 218234.168 91.385 88.600
27 28012.716 95.166 215142.502 91.369 88.601
28 27136.307 95.194 212505.200 91.413 88.660
29 26781.810 95.206 210085.084 91.404 88.638
30 26448.669 95.202 207351.591 91.388 88.635
Added tags to vectors: https://gist.github.com/sskorol/2dcc110a58e932810ec55671e849a979
Itn Tag Loss Tag % Dep Loss UAS LAS
--- --------- -------- --------- ------ ------
1 148579.293 93.289 602149.213 87.441 82.868
2 86696.120 94.096 461888.448 89.006 85.180
3 74762.131 94.472 415305.713 89.697 86.195
4 66960.488 94.667 384642.778 90.292 86.975
5 61450.382 94.830 360781.163 90.503 87.291
6 56722.733 94.955 343727.112 90.727 87.587
7 53009.846 95.001 328229.325 90.834 87.801
8 50004.985 95.054 314413.512 90.970 87.975
9 47342.091 95.083 303663.900 91.089 88.138
10 44426.744 95.161 294157.662 91.210 88.281
11 42814.691 95.206 284806.755 91.214 88.331
12 40708.372 95.219 277395.015 91.212 88.381
13 39001.517 95.294 271132.501 91.200 88.379
14 37216.418 95.285 263681.832 91.277 88.496
15 36106.172 95.287 256903.847 91.345 88.589
16 34565.183 95.275 252155.354 91.396 88.636
17 33421.656 95.259 246689.826 91.434 88.681
18 32859.640 95.306 241977.253 91.496 88.754
19 31423.111 95.321 236369.869 91.491 88.730
20 30424.943 95.317 232551.682 91.479 88.721
21 29464.565 95.340 227281.395 91.502 88.758
22 28680.611 95.350 224486.824 91.480 88.760
23 28129.433 95.344 220624.970 91.505 88.788
24 27250.755 95.371 218851.535 91.568 88.834
25 26648.756 95.376 214631.149 91.613 88.866
26 25724.630 95.381 211835.240 91.641 88.906
27 25347.578 95.392 208937.050 91.579 88.869
28 24685.303 95.375 203972.069 91.610 88.912
29 24342.255 95.418 202738.656 91.621 88.927
30 24096.019 95.385 200364.226 91.641 88.959
Training and testing on nerus (?) dataset is a special case for a news media domain. As some people pointed out the NER e.g. won't recognize lower case entities. It would be interesting to see if adding some noise (lowercasing, misspelling) would lead to recognising more entities, say in a casual chat.
@sbushmanov experiments say that while this is improving quality on lowercase-entries, this is decreasing quality on well-written texts. so if you have mixed cases, you need a mixed-case model for such texts, but better to use a good model for properly spelled/cased texts.
vec = None:
vec = paragidms_from_navec, width=96: